JP2012203853A - Image processing device, image processing method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a new and improved image processing device or the like, for detecting horizontal parallax with higher robustness and accuracy.SOLUTION: The image processing device includes a horizontal parallax detecting part that detects horizontal parallax of a reference pixel, based on the result of comparing a reference pixel constituting a reference image with a first pixel to be referred located in the same height position as the reference pixel and a second pixel to be referred located in a height position different from the first pixel to be referred constituting an image to be referred, in the reference image and the image to be referred in which the same subject is drawn in different horizontal positions.

Description

  The present invention relates to an image processing apparatus, an image processing method, and a program.

  There is known an autostereoscopic display device capable of displaying an image in a stereoscopic manner even without glasses dedicated to stereoscopic vision. The autostereoscopic display device acquires a plurality of images in which the same subject is drawn at different horizontal positions. Then, the autostereoscopic display device compares subject images, which are portions where the subject is drawn, among these images, and detects a horizontal position shift between the subject images, that is, horizontal parallax. Then, the autostereoscopic display device generates a plurality of multi-view images based on the detected horizontal parallax and the acquired image, and stereoscopically displays these multi-view images. Local matching described in Patent Document 1 and global matching described in Patent Document 2 are known as methods for detecting the horizontal parallax by the autostereoscopic display device.

JP 2006-164297 A Patent No. 4410007

  However, local matching has a problem that the robustness and accuracy of parallax detection are low. On the other hand, in the global matching, when the quality of the image acquired by the autostereoscopic display device is high, that is, when the vertical positions of the subject images are not shifted (there is no vertical shift), the parallax detection has robustness. Although it was high, there was a problem that the robustness and accuracy of the parallax detection greatly deteriorated when there was a shift in the vertical position between the subject images (there was a vertical shift).

  Therefore, the present invention has been made in view of the above problems, and an object of the present invention is a new and improved detection capable of detecting horizontal parallax with higher robustness and accuracy than conventional techniques. An image processing apparatus, an image processing method, and a program are provided.

  In order to solve the above-described problem, according to a certain aspect of the present invention, a reference pixel and a reference image that constitute a reference image among reference images and reference images in which the same subject is drawn at different horizontal positions are configured. Of the pixels, the first reference pixel existing at the same height position as the reference pixel and the second reference pixel having a height position different from the first reference pixel are compared, and based on the result of the comparison An image processing apparatus including a horizontal parallax detection unit that detects the horizontal parallax of the reference pixel is provided.

  Here, the horizontal parallax detection unit includes a first evaluation value that evaluates a difference between a feature amount in the reference region including the reference pixel and a feature amount in the first reference region including the first reference pixel; A second evaluation value that evaluates a difference between the feature amount in the reference region and the feature amount in the second reference region including the second reference pixel, and calculates the first evaluation value and the second evaluation value; The horizontal parallax of the reference pixel may be detected based on the evaluation value.

  The horizontal parallax detection unit includes a correction execution unit that corrects the standard image and the reference image before the horizontal parallax is detected, and a correction reliability calculation unit that calculates the reliability of the correction performed by the correction execution unit. The first evaluation value and the second evaluation value may be calculated based on the reliability calculated by the correction reliability calculation unit.

  Further, the horizontal parallax detection unit calculates a change amount per unit time of the feature amount of the reference pixel, and calculates the first evaluation value and the second evaluation value based on the calculated change amount. Also good.

  Further, the horizontal parallax detection unit generates a mosaic image by reducing the resolution of the reference image, calculates a difference between the feature amount of the reference pixel in the reference image and the feature amount of the reference pixel in the mosaic image, and calculates The first evaluation value and the second evaluation value may be calculated based on the difference.

  Further, the horizontal parallax detection unit has the horizontal position of the reference pixel and the horizontal parallax of the reference pixel as components of the node, and among the first evaluation value and the second evaluation value, the first evaluation value and the second evaluation value A DP map is generated in which the evaluation value with the smaller absolute value of the difference evaluated by the evaluation value becomes the score of the node, and the difference between the feature amounts of the reference pixel and the pixels in the peripheral area of the reference pixel and the reference image are configured Pixels from the start point to the node based on the difference in feature amount between the reference corresponding pixel existing at the same position as the reference pixel and the pixel in the peripheral region of the reference corresponding pixel and the node score The accumulated cost may be calculated, the shortest path that minimizes the accumulated cost from the start point to the end point may be calculated, and the horizontal parallax of the reference pixel may be detected based on the calculated shortest path.

  Further, the difference in height position between the second reference pixel and the first reference pixel may be limited within a predetermined range.

  In addition, the second horizontal parallax detection unit that detects the horizontal parallax of the reference pixel by a detection method different from the first horizontal parallax detection unit that is the horizontal parallax detection unit, and the reference pixel detected by the first horizontal parallax detection unit Disparity for calculating the reliability of the horizontal parallax of the reference pixel and the horizontal parallax of the reference pixel detected by the second horizontal parallax detection unit, and generating a disparity map indicating the high-reliability horizontal parallax for each reference pixel And a map integration unit.

  Further, the disparity map integration unit evaluates the reliability of the horizontal parallax of the reference pixel detected by the first horizontal parallax detection unit and the reliability of the horizontal parallax of the reference pixel detected by the second horizontal parallax detection unit, respectively. You may make it calculate by.

  According to another aspect of the present invention, among the standard image and the reference image in which the same subject is drawn at different horizontal positions, the standard pixel constituting the standard image and the standard pixel among the pixels constituting the reference image The first reference pixel present at the same height position and the second reference pixel having a height position different from that of the first reference pixel are compared, and the horizontal parallax of the reference pixel is determined based on the comparison result. An image processing method is provided that includes a detecting step.

  According to another aspect of the present invention, among a standard image and a reference image in which the same subject is drawn at different horizontal positions on a computer, among a standard pixel that constitutes the standard image and a pixel that constitutes the reference image, A first reference pixel that is present at the same height as the reference pixel and a second reference pixel that is different in height from the first reference pixel are compared, and based on the comparison result, the reference pixel A program for realizing a horizontal parallax detection function for detecting horizontal parallax is provided.

  As described above, the present invention considers the second reference pixel at a height different from the reference pixel in addition to the first reference pixel at the same height as the reference pixel when detecting the horizontal parallax of the reference pixel. Therefore, it is possible to detect horizontal parallax that is strong against geometrical deviation (vertical deviation). As a result, the present invention can detect horizontal parallax with higher robustness and higher system than before.

It is a flowchart which shows the outline | summary of the process by an autostereoscopic display apparatus. It is explanatory drawing which shows the color shift of input images. It is explanatory drawing which shows the geometric shift of input images. It is explanatory drawing which shows a mode that a parallax map (disparity map) and a multi-viewpoint image are produced | generated. 1 is a block diagram illustrating a configuration of an image processing apparatus according to an embodiment of the present invention. It is a block diagram which shows the structure of a color calibration part. It is a block diagram which shows the structure of a histogram matching part. It is explanatory drawing which shows a mode that an input image is divided | segmented into a some block. It is a color histogram generated for each block. It is a DP map for color correspondence detection used for color histogram matching. It is a block diagram which shows the structure of a linear color parameter fitting part. It is explanatory drawing which compares and shows the gamma curve in each of the image which performed color calibration, and the image which has not performed color calibration. It is a block diagram which shows the structure of a non-linear color parameter fitting part. It is a block diagram which shows the structure of an evaluation part. It is explanatory drawing which compares and compares the multiview image produced | generated after performing color calibration, and the multiview image produced | generated without performing color calibration. It is a block diagram which shows the structure of a geometric calibration part. It is explanatory drawing which shows the input image in which the feature-value matching process was performed. It is a block diagram which shows the structure of a corresponding point refinement part. It is a histogram which shows the frequency distribution of the vector in each block. It is explanatory drawing which shows the mode of a mean shift (Mean Shift) process. It is a block diagram which shows the structure of a linear geometric parameter fitting part. It is a block diagram which shows the structure of an initialization part. It is a graph for determining the weight which shows how much time restrictions apply to F matrix. It is a block diagram which shows the structure of a linear geometric parameter calculation part. It is a block diagram which shows the structure of a nonlinear geometric parameter fitting part. It is a graph which shows a mode that the parameter of F matrix which is not restrained temporally, and the parameter of F matrix restrained temporally change according to progress of time. It is a graph which shows a mode that F matrix converges with progress of time. It is a block diagram which shows the other example of a linear geometric parameter fitting part. It is a block diagram which shows the structure of a projection component calculation part. It is a block diagram which shows the structure of a rotation component calculation part. It is a block diagram which shows the structure of a translation component calculation part. It is a block diagram which shows the other example of a linear geometric parameter fitting part. It is an update table used when updating F matrix. It is a graph which shows a mode that F matrix updated by the update table converges according to progress of time. It is a graph which shows the correspondence of the component of F matrix which performed geometric calibration, and an error value. It is explanatory drawing which shows the mode of the geometric calibration by image rubbing. It is a block diagram which shows the structure of a parallax detection part. It is a block diagram which shows the structure of a global matching part. It is a block diagram which shows the structure of a motion description part. It is a block diagram which shows the structure of an anchor vector construction part. It is a local histogram which shows the frequency of parallax. It is a block diagram which shows the structure of a route construction part. It is a DP map used when performing parallax matching. It is explanatory drawing which shows a reliability map. It is explanatory drawing which shows the multiview image produced | generated based on various parallax maps. It is a flowchart which shows the procedure of color calibration. It is a flowchart which shows the procedure of geometric calibration. It is a flowchart which shows the procedure of parallax detection.

  Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function structure, duplication description is abbreviate | omitted by attaching | subjecting the same code | symbol.

The description will be made in the following order.
1. 1. Overview of processing performed by the autostereoscopic display device 2. Configuration of image processing apparatus 2-1. Configuration of color calibration unit 2-2. Configuration of geometric calibration unit 2-3. 2. Configuration of parallax detection unit Processing performed by the image processing apparatus

<1. Overview of processing performed by autostereoscopic display device>
The inventor of the present application has earnestly researched an autostereoscopic display device capable of stereoscopic display of an image without glasses dedicated to stereoscopic viewing, and invented the image processing apparatus according to the present embodiment in the process. It came to do. Here, the stereoscopic display means that an image is stereoscopically displayed by causing binocular parallax to be generated by the viewer.

  First, processing performed by the autostereoscopic display device including the image processing device will be described with reference to the flowchart shown in FIG.

In step S1, autostereoscopic display apparatus acquires an input image _V L, _{V R} of the current frame. Figure 2 shows an example of the input image _V L, _{V R} in FIG. In the present embodiment, the input image V _L, the pixel of the upper left point of the V _R as the origin, x-axis in the horizontal direction, the vertical direction is y-axis. The right direction is the positive direction of the x axis, and the downward direction is the positive direction of the y axis. Each pixel has coordinate information (x, y) and color information (luminance (brightness), saturation, hue). Hereinafter, the input image V _L also referred pixels on _(and the input image V _L to the image obtained by performing the calibration) and "left coordinate point", the calibration input image V _{R (and} the input image V _R The pixel on the image obtained by performing the above is also referred to as a “right coordinate point”.

Figure 2, as shown in FIG. 3, are drawn to the same subject (eg sea, fish, penguin) of different horizontal position (x-coordinate) of the input image V _L, V _R.

Input image _V L, _{V R} shown in Figure 2, it has occurred colors displaced from each other. That is, drawn in a different color in the same subject as the input image V _L and the input image V _R. For example, the subject image V _L 1 and the subject image V _R 1 both indicate the same sea, but have different colors. On the other hand, the input image _V L, _{V R} shown in Figure 3, has occurred geometric displaced from each other. That is, the same subject is drawn at different height positions (y coordinates). For example, the subject image V _L 2 and the subject image V _R 2 both show the same penguin, but the y coordinate of the subject image V _L 2 and the y coordinate of the subject image V _R 2 are different from each other. In FIG. 3, straight lines L <b> 1 and L <b> 2 are drawn in order to facilitate understanding of the geometric misalignment.

Therefore, the autostereoscopic display device reduces these deviations by performing color calibration and geometric calibration. That is, the color calibration, the input image V _L, a process of correcting a color shift of V _R, geometric calibration is a process for correcting the geometric shift of the input image V _L, V _R.

In step S2, the naked-eye stereoscopic display device based on color calibration and geometric calibrated input image V _L, calibrated image V _L-2 is _{V R,} V _R-2, performs a parallax detection . The state of parallax detection is shown in FIG.

As shown in FIG. 4, the autostereoscopic display device includes a horizontal parallax d1 (between the coordinate point P _L- 21 on the calibrated image V _L-2 and each point on the epipolar line EP _R-2 1. calculating the x difference between coordinates), the optimum ones of these, i.e. the horizontal parallax d1 between the coordinate point P _{R-2 1} and the coordinate point P _{L-2 1} horizontal parallax d1. Here, the epipolar line EP _R- 21 is a straight line drawn on the calibrated image VR _-2 , having the same y coordinate as the coordinate point P _L- 21 and extending in the horizontal direction. Such a search is possible because the autostereoscopic display device performs geometric calibration in advance, and the geometric misalignment between the calibrated image V _L-2 and the calibrated image V _R-2 is reduced. Because.

The autostereoscopic display device generates a parallax map (disparity map) DM by calculating the horizontal parallax d1 for all the pixels on the calibrated image V _L-2 . In the parallax map DM, the horizontal parallax d1 is described for all the pixels on the calibrated image _VL-2 . In FIG. 4, the degree of the horizontal parallax d1 is indicated by shades of hatching.

In step S3, the autostereoscopic display device generates a plurality of multi-viewpoint images V _V based on the parallax map DM and the calibrated images V _L-2 and V _R-2 . For example, the multi-viewpoint image V _V illustrated in FIG. 4 is an image that interpolates between the calibrated image V _L-2 and the calibrated image VR _-2 . Thus, the pixel _{P V} 1 showing the same point as the point on the object indicated by the pixel _{P L-2} 1 is present between the pixels _{P L-2} 1 and the pixel _{P R-2} 1.

Here, each of multi-viewpoint images V _V is an image to be stereoscopically displayed by autostereoscopic display device, corresponding to different viewpoints (viewer's eye position). That is, the multi-viewpoint images V _V eyes of a viewer viewing will vary depending on the eye position of the viewer. For example, since the viewer's right eye and left eye are present at different positions, the different multi-view images V _V are visually recognized. Thus, a viewer, a multi-viewpoint image V _V can stereoscopically view. Further, by the viewer moves, they change viewer's perspective, if there is a multi-viewpoint image V _V corresponding to the viewpoint, the viewer can stereoscopically multi-viewpoint images V _V. Thus, as the number of multi-viewpoint images V _V is large, the viewer has the multi-viewpoint images V _V can be stereoscopically viewed on more positions. Further, as the number of multi-viewpoint images V _V increases, the reverse viewing, that is, the phenomenon that the viewer visually recognizes the multi-viewpoint images V _V that should be visually recognized with the right eye is less likely to occur. Further, by creating multiple multi-viewpoint images V _V, it is possible to express motion parallax.

In step S4, the autostereoscopic display device performs fallback. This process is schematically a process of re-corrected according to the multi-viewpoint images V _V of the agreement. In step S5, autostereoscopic display device displays stereoscopic each multi-viewpoint images V _V.

<2. Configuration of Image Processing Device>
Next, the configuration of the image processing apparatus 1 according to the present embodiment will be described with reference to the drawings. As shown in FIG. 5, the image processing apparatus 1 includes a color calibration unit 2, a geometric calibration unit 3, and a parallax detection unit 4. Color calibration unit 2 performs color calibration as described above with respect to the input image V _L, V _R of the current frame, the geometric calibration unit 3 is an input image already color calibration V _L, V _R color The above _- described geometric calibration is performed on the calibrated images V _L-1 and V _R-1 . The parallax detection unit 4 performs the above _- described parallax detection on the calibrated images V _L-2 and VR _-2 . Therefore, the image processing apparatus 1 performs the above-described steps S1 to S2 among the processes performed by the autostereoscopic display device.

<2-1. Configuration of color calibration unit>
Next, the configuration of the color calibration unit 2 will be described with reference to FIG. As shown in FIG. 6, the color calibration unit 2 includes a histogram matching unit 201, a linear color parameter fitting unit 202, a feature amount matching unit 203, a nonlinear color parameter fitting unit 204, an evaluation unit 205, and a color calibration. An execution unit 206.

[Configuration of histogram matching unit]
As shown in FIG. 7, the histogram matching unit 201 includes a block division unit 207, a histogram generation unit 208, a DP (Dynamic Programming) matching unit 209, and a color pair calculation unit 210.

As shown in FIG. 8A, the block division unit 207 divides the input image V _L into eight pixel blocks B _L−1 . Here, each pixel block _BL-1 is rectangular and all have the same size. Further, the block dividing unit 207 extracts the pixel that is the center point of each pixel block B _L-1 , and sets the coordinates of the extracted pixel as the coordinates of each pixel block B _L-1 .

Similarly, the block dividing unit 207, as shown in FIG. 8 (b), divides the input image _{V R} into eight pixel blocks _{B R-1.} Here, each pixel block _BR-1 has the same size and shape as the pixel block _BL-1 . The block division unit 207 extracts a pixel which is the center point of the pixel blocks B _R-1, the coordinates of the extracted pixels and the pixel blocks B _R-1 coordinates. Therefore, each pixel block _BL-1 includes a pixel block _BR-1 having the same coordinates. In other words, the pixel block _BL-1 and the pixel block _BR-1 correspond one-to-one. Therefore, the block division unit 207 associates the pixel block _BL-1 with the pixel block BR _-1 . Note that the number of pixel blocks B _L-1 and B _R-1 is not limited to eight.

Then, the block dividing unit 207 reduces the gradation in each pixel block B _L-1 and _BR-1 to 64 gradations. Note that the gradation in each of the pixel blocks B _L-1 and B _R-1 is not limited to 64 gradations. Original input image V _L, but the tone is not particularly limited also in V _R, for example, 1024 gradations. Then, the block dividing unit 207 outputs block information regarding each pixel block B _L−1 and _BR−1 to the histogram generating unit 208.

The histogram generation unit 208 performs the following processing based on the block information. That is, the histogram generation unit 208 generates a color histogram CH _L as shown in FIG. 9A for each pixel block B _L-1 . The color histogram CH _L indicates the luminance distribution in the pixel block B _L-1 , that is, the correspondence between the luminance and the number (frequency) of pixels having the luminance. Here, the luminance is given as an integer value from 0 to 63. Similarly, the histogram generation unit 208 generates a color histogram CH _R as shown in FIG. 9B for each pixel block _BR-1 . The color histogram CH _R indicates the correspondence between the luminance distribution in the pixel block _BR-1 , that is, the luminance and the number (frequency) of pixels having the luminance. Since the pixel block B _L-1 and the pixel block B _R-1 have a one-to-one correspondence, the color histogram CH _L and the color histogram CH _R have a one-to-one correspondence. Therefore, the histogram generation unit 208 associates the color histogram CH _L and color histogram CH _R. Then, the histogram generation unit 208 generates color histogram information regarding the color histograms CH _L and CH _R and outputs the color histogram information to the DP matching unit 209.

DP matching unit 209 (matching unit) on the basis of the color histogram information, a pixel block _{B L-1,} the Kumigoto the pixel blocks _{B R-1} corresponding to the pixel block _{B L-1,} the following process Do. That is, the DP matching unit 209 first generates a color correspondence detection DP map as shown in FIG. The vertical axis of the color correspondence detection DP map indicates the frequency for each luminance in the pixel block _BL-1 . For example, FL (n _L ) indicates the frequency of luminance n _L (n _L is an integer of 0 to 63). On the other hand, the horizontal axis of the DP map for color correspondence detection indicates the frequency for each luminance in the pixel block _BR-1 . For example, FL (n _R ) indicates the frequency of luminance n _R (n _R is an integer of 0 to 63). A node indicating the frequency FL (n _L ) and the frequency FL (n _R ) is represented by P (n _L , n _R ).

Then, the DP matching unit 209 uses the node P (0, 0) as the start point, the node P (63, 63) as the end point, and the accumulated cost from the start point to the node P (n _L , n _R ) as follows: Defined in

_{Here, DFI (n L, n R} ) 0 is the cumulative cost of time leading to node _P through a path _{PA C 0 (n L, n} R), DFI (n L, n R) 1 is node _{P (n} L, n _R) through a path PA _C 1 is a cumulative cost when leading _{to, DFI (n L, n R} ) 2 , the node P _(n L through the path PA _C 2, n _R ) is the accumulated cost. DFI (n _L , n _R −1) is an accumulated cost from the starting point to the node P (n _L , n _R −1). DFI (n _L −1, n _R ) is an accumulated cost from the start point to the node P (n _L −1, n _R ). DFI (n _L −1, n _R −1) is an accumulated cost from the start point to the node P (n _L −1, n _R −1).

The DP matching unit 209 calculates the accumulated cost from the start point to each node including the end point, and by tracing back the route having the minimum accumulated cost from the end point to the start point, the shortest route, that is, from the start point to the end point. The route with the minimum accumulated cost is calculated. The nodes on this shortest path show similar brightness pairs. For example, when the shortest path passes through the node P (n _L , n _R ), the luminance n _L and the luminance n _R are similar. The DP matching unit 209 generates the shortest path information regarding the shortest path calculated for each set of the pixel blocks B _L-1 and B _R-1 and outputs the shortest path information to the color pair calculation unit 210.

Color Pair calculator 210, based on the shortest path information, calculates the color pairs for each set of pixel blocks _{B L-1, B R-} 1. That is, the color pair calculation unit 210 calculates a luminance pair indicated by a node on the shortest path, that is, a luminance pair similar to each other as a color pair. The color pair calculation unit 210 calculates about 250 color pairs per pixel block _BL-1, and calculates about 2000 color pairs in total. The color pair calculation unit 210 generates color pair information regarding the calculated color pair, and outputs the color pair information to the linear color parameter fitting unit 202 and the non-linear color parameter fitting unit 204.

Accordingly, histogram matching unit 201, the pixel block _{B L-1, B R-} 1 in the input image _V L gradation of, since drop than the gradation of _{V R,} histogram _CH L, the amount of data, such as CH _R Can be reduced. Note that by gradation falls, the input image V _L, since a plurality of types of luminance of V _R is to be classified into one of the luminance, it is assumed that the accuracy of color calibration is reduced. Therefore, the histogram matching unit 201 compensates for the decrease in accuracy by calculating a color pair with a plurality of pixel blocks B _L-1 and B _R-1 , that is, by increasing the number of color pairs.

[Configuration of linear color parameter fitting unit]
As shown in FIG. 11, the linear color parameter fitting unit 202 includes an addition matrix calculation unit 211, a coefficient calculation unit 212, a filter unit 213, and a coefficient holding unit 214.

Linear color parameter fitting portion 202, on the basis of the color pair information, is calculated by linear operations the model equation showing the relationship between the luminance of the input image V _R of the input image V _L. First, this model formula will be described.

Input image V _L, the gamma value of V _R is assumed to be near 1, the luminance of the input image V _L, using the luminance of the input image V _R is expressed as follows.

Here, L is the luminance of the input image V _L, R is the luminance of the input image V _R. α is the gain of the autostereoscopic display device, and γ is the gamma value of the autostereoscopic display device. That is, the input image V _L, the brightness of the V _R is constrained by the gain and gamma value. In this way, a dull image can be obtained. When Expression (4) is Taylor-expanded, the following Expression (6) is obtained.

Here, w _i (i is an integer of 0 or more) is a coefficient, that is, a color correction coefficient. Equation (6) is a model equation. Note that equation (6) below by adding an offset term w _a formula (7) can be accurately follow the change in the luminance of the input image V _L, V _R than equation (6). Therefore, the linear color parameter fitting unit 202 uses the formula (6) or the formula (7) as a model formula (that is, an approximate formula).

Here, coef (w) represents color correction coefficients w _i and w _a .

  Next, processing performed by each component will be described. The addition matrix calculation unit 211 calculates a matrix represented by the following expression (8) for all color pairs.

  Here, n1 is an arbitrary integer equal to or greater than 2, and the larger the value of n1, the greater the number of calculated coefficients, and the more accurate the model formula is. The value of n1 is determined according to required accuracy, processing speed, and the like, and is 2 to 5, for example.

Next, the addition matrix calculation unit 211 calculates an addition matrix M ₁ represented by the following equation (9) by adding all the calculated matrices.

The addition matrix calculation unit 211 generates addition matrix information regarding the calculated addition matrix M ₁ and outputs the addition matrix information to the coefficient calculation unit 212 together with the color pair information.

  The coefficient calculation unit 212 (model formula calculation unit) calculates a matrix represented by the following formula (10) for all color pairs based on the color pair information.

  The coefficient calculation unit 212 calculates a coefficient calculation matrix A represented by the following formula (11) by adding all the calculated matrices.

Next, the coefficient calculation unit 212 calculates the first initial value coef (w) ₁ of the color correction coefficients wi and wa based on the following equation (12).

The coefficient calculation unit 212 generates first initial value information regarding the calculated first initial value coef (w) ₁ and outputs the first initial value information to the filter unit 213.

The filter unit 213 includes first initial value information given from the coefficient calculation unit 212, linear correction coefficient information on the previous frame (one frame before the current frame) given from the coefficient holding unit 214, and Based on the equation (13), the second initial value coef (w) ₂ of the color correction coefficients wi and wa is calculated.

Here, oldcoef (linear) is the linear correction coefficient of the previous frame, that is, the color correction coefficients wi and wa of the previous frame, and w _b is the linear correction coefficient oldcoef (linear) of the previous frame as the second initial value coef (w ) A weight indicating how much to act on ₂ . The larger the value of the weight w _b (the greater the weight), the larger the linear correction coefficient oldcoef (linear) of the previous frame acts on the second initial value coef (w) ₂ . The value of the weight w _b is, for example, 0.5 to 0.9.

Next, the filter unit 213 causes the IIR filter to act on the second initial value coef (w) ₂ . The filter unit 213 uses the color correction coefficients wi and wa obtained thereby as the linear correction coefficient coef (linear), linear correction coefficient information regarding the linear correction coefficient coef (linear) as a coefficient holding unit 214, and a nonlinear color parameter fitting unit 204. And output to the evaluation unit 205 and the color calibration execution unit 206.

The coefficient holding unit 214 holds the linear correction coefficient information given from the filter unit 213, and when the second initial value coef (w) ₂ of the next frame is calculated, It outputs to the filter part 213 as linear correction coefficient information of a flame | frame.

12, the input image _V L corrected by the linear correction coefficient coef (linear), and gamma curve L1 of _{V R,} the input image _V L when not subjected to correction, gamma value and a gamma curve L2 of _{V R} Each is shown in comparison. That is, γ = 0.6 in FIG. 12A, γ = 0.8 in FIG. 12B, γ = 1.2 in FIG. 12C, and γ = 1.4 in FIG. It has become. In FIG. 12, equation (6) is a model equation, and n1 = 2. According to FIG 12, color calibration unit 2, using the linear correction coefficient coef (linear) by performing a color calibration, the input image V _L, while substantially maintaining the brightness and vividness of V _R The brightness can be corrected.

[Configuration of feature matching unit]
Feature matching unit 203 shown in FIG. 6, by matching the brightness of the input image V _L, V _R of the current frame, and extracts the color pair. Specifically, feature matching unit 203, the input image _V L, from _{V R,} the predetermined value (e.g., 32 to 1024) extracts a feature point having a brightness above. Next, the feature amount matching unit 203 performs the following processing for each feature point of the input image _VL . That is, the feature amount matching unit 203 exists in the vicinity of the feature point of the input image V _L (that is, the absolute value of the x coordinate difference is 64 to 256 or less, and the absolute value of the y coordinate difference is 8 to 8). 64 of the following become) feature points of the input image V _R, the luminance is most similar (i.e., the smallest absolute value of the difference in brightness) to extract the feature points, the feature point of the input image V _L and the characteristic point Are feature point pairs, that is, color pairs. Next, the feature amount matching unit 203 generates color pair information regarding the extracted color pair, and outputs the color pair information to the nonlinear color parameter fitting unit 204 and the evaluation unit 205.

[Configuration of nonlinear color parameter fitting unit]
The non-linear color parameter fitting unit 204 performs color correction coefficient wi based on the color pair information given from the histogram matching unit 201 and the feature amount matching unit 203 and the linear correction coefficient information given from the linear color parameter fitting unit 202. , Wa are calculated by nonlinear calculation.

  Specifically, the non-linear color parameter fitting unit 204 includes a Jacobian matrix calculating unit 215, a Hessian matrix calculating unit 216, and a coefficient updating unit 217, as shown in FIG.

Based on the linear correction coefficient information given from the linear color parameter fitting unit 202 and the color pair information given from the histogram matching unit 201 and the feature amount matching unit 203, the Jacobian matrix calculation unit 215 has the following formula (14 ) to calculate the Jacobian matrix J ₁ represented by.

Here, L is the luminance of the input image _{V L,} R is the luminance of the input image _{V R,} coef (w) denotes a color correction coefficient wi, the wa. f _e1 (L, R, coef (w)) is a so-called error function, and is represented by the following equation (15). f _{e10 to} f _e1m are obtained by substituting different color pair information into the error function f _e1 (L, R, coef (w)).

_Here, g 1 (R, coef (w)), as shown in Equation (7), brightness and color correction coefficients of the input image _{V R} the luminance of the input image _{V L} wi, is an equation approximated by wa Therefore, the value of (L−g ₁ (R, coef (w))) indicates an error between the actual value of the input image V _L and the approximate value. RobustFunc is a function that normalizes the error to a value in the range of 0 to 1, and is represented by the following equation (16), for example.

Here, w _c is a coefficient (weight), for example, 0.1 to 8.0. In Equation (15), the error is (L−g ₁ (R, coef (w))), but the error is another value, for example, (g ₁ (L, coef (w)) − g ₁ (R, coef (w))). _Here, g 1 (L, coef (w)), the luminance and color correction coefficients of the input image _{V L} the luminance of the input image _{V R} wi, a value approximated by wa, the following equation (7-1) And (7-2).

Then, the Jacobian matrix calculating unit 215 generates Jacobian matrix information related to the calculated Jacobian matrix J ₁ and outputs the generated Jacobian matrix information to the Hessian matrix calculating unit 216.

Further, when the Jacobian matrix calculation unit 215 is given nonlinear correction coefficient information from a coefficient updating unit 216 described later, the Jacobian matrix calculation unit 215 is based on the nonlinear correction coefficient coef (nonlinear) indicated by the nonlinear correction coefficient information and the color pair information. Te to calculate the Jacobian matrix J ₁ described above. Then, the Jacobian matrix calculating unit 215 generates Jacobian matrix information related to the calculated Jacobian matrix J ₁ and outputs the generated Jacobian matrix information to the Hessian matrix calculating unit 216.

The Hessian matrix calculation unit 216 calculates the Hessian matrix H ₁ represented by the following equation (17) based on the information given from the Jacobian matrix calculation unit 216.

Next, the Hessian matrix calculation unit 216 generates Hessian matrix information regarding the calculated Hessian matrix H _1, and outputs it to the coefficient update unit 217 together with the information given from the Jacobian matrix calculation unit 215.

The coefficient updating unit 217 calculates the color correction coefficients w _i and w _a as nonlinear correction coefficients coef (nonlinear) based on the information given from the Hessian matrix calculation unit 216 and the following equation (18).

Here, I is a unit matrix. w _d is a coefficient (weight), for example, and from 1.0 to 0.0.

  Next, the coefficient updating unit 217 generates nonlinear correction coefficient information regarding the calculated nonlinear correction coefficient coef (nonlinear), and determines whether the nonlinear correction coefficient coef (nonlinear) has converged to a certain value. When it is determined that the nonlinear correction coefficient coef (nonlinear) has converged, the coefficient update unit 217 outputs nonlinear correction coefficient information to the evaluation unit 205 and the color calibration execution unit 206, and the nonlinear correction coefficient coef (nonlinear) If it is determined that it has not converged, the nonlinear correction coefficient information is output to the Jacobian matrix calculation unit 215. Thereafter, the non-linear correction coefficient coef (nonlinear) is recalculated. Therefore, the non-linear color parameter fitting unit 204 calculates the non-linear correction coefficient coef (nonlinear) using the linear correction coefficient coef (linear) as an initial value, and the non-linear correction coefficient until the non-linear correction coefficient coef (nonlinear) converges to a constant value. Repeat the calculation of coef (nonlinear).

[Configuration of evaluation section]
As shown in FIG. 14, the evaluation unit 14 (correction reliability calculation unit) includes a color conversion unit 218, a histogram base reliability calculation unit 219, a feature point base reliability calculation unit 220, and a reliability map generation unit 221. With.

Color conversion unit 218 acquires an input image _V L, _{V R} of the current frame. Color conversion unit 218, the input image V _L, and V _R, and the linear correction coefficient information given from the linear color parameter fitting portion 202, on the basis to correct the brightness of the input image V _R. Specifically, the color conversion unit 218, the luminance of the input image _{V R,} by substituting the _g 1 (R, coef) of the linear correction coefficient coef (linear) and Equation _{(7), g} 1 (R , Coef). Calculated values, it indicates luminance of the input image V _L corresponding to the brightness of the input image V _R, the color conversion unit 218, this value and the brightness of the new input image V _R. Thus, the difference between the luminance of the input image V _R of the input image V _L is reduced. Color conversion unit 218 outputs the histogram-based reliability calculation unit 219 and the input image _{V R} and the corrected input image _{V L} linear color calibrated image _{V L-1a,} as _{V R-1a.}

The histogram-based reliability calculation unit 219 divides the linear color calibrated images V _L-1a and V _R-1a into eight pixel blocks B _L-1 and _BR-1 . Here, the pixel blocks B _L-1 and B _R-1 are the same as those generated by the histogram matching unit 201. Then, the histogram-based reliability calculation unit 219 generates a corrected color histogram indicating the correspondence between the luminance and the number (frequency) of pixels having the luminance for each of the pixel blocks B _L-1 and B _R-1. To do.

Then, the histogram-based reliability calculation unit 219 uses the normalized cross-correlation (NCC) for the similarity between the corrected color histogram of the pixel block _{BL-1 and} the corrected color histogram of the pixel block _BR-1. To calculate. Note that the method for calculating the similarity is not particularly limited, and for example, an intersection, Bhattacharya, or the like may be used.

The histogram-based reliability calculation unit 219 calculates the similarity for each pixel block _BL-1 , and substitutes these similarities into the error of the above-described formula (16), for example, to reduce the similarity between 0 and 1. Normalize to a value within the range. Then, the histogram base reliability calculation unit 219 sets the normalized value as the reliability of the linear correction coefficient coef (linear). Then, the histogram-based reliability calculation unit 219 generates linear correction coefficient reliability information regarding the calculated reliability of each pixel block _BL-1 , and outputs the linear correction coefficient reliability information to the reliability map generation unit 221.

The feature point base reliability calculation unit 220 classifies the feature points in the input image V _L into each of the eight pixel blocks B _L−1 based on the color pair information given from the feature amount matching unit 203. Here, the pixel block _BL-1 is the same as that generated by the histogram matching unit 201. Then, the feature point base reliability calculation unit 220 calculates the reliability of the nonlinear correction coefficient coef (nonlinear) for each pixel block B _L−1 based on the nonlinear correction coefficient information given from the nonlinear color parameter fitting unit 204. To do.

That is, the feature point base reliability calculation unit 220 calculates an error for each feature point in the pixel block B _L-1 using the error function f _e1 (L, R, coef) described above. That is, the feature point-based reliability calculation unit 220 calculates by substituting the value of the color pair information for L and R of f _e1 (L, R, coef) and the value of the nonlinear correction coefficient coef (nonlinear) for coef. Thus, the error is calculated. And the feature point base reliability calculation part 220 extracts the feature point from which an error is below a predetermined value (for example, 0.5-8.0), and calculates the average value of an error about the extracted feature point. Then, the feature point base reliability calculation unit 220 sets the calculated average value as the reliability of the nonlinear correction coefficient coef (nonlinear). Note that the feature point-based reliability calculation unit 220 sets the reliability to zero when the number of feature points whose error is equal to or less than a predetermined value is equal to or less than a predetermined value (for example, 0 to 8). Thereby, the feature point base reliability calculation unit 220 calculates the reliability of the nonlinear correction coefficient coef (nonlinear) for each pixel block _BL-1 . The feature point base reliability calculation unit 220 generates nonlinear correction coefficient reliability information related to the calculated reliability of each pixel block _BL-1 , and outputs it to the reliability map generation unit 221.

The reliability map generation unit 221 generates a color calibration reliability map based on information given from the histogram base reliability calculation unit 219 and the feature point base reliability calculation unit 220. Specifically, the reliability map generation unit 221 compares the reliability of the linear correction coefficient coef (linear) and the reliability of the nonlinear correction coefficient coef (nonlinear) for each pixel block _BL-1 , and calculates the larger one. Is the reliability of the pixel block _BL-1 . In addition, the reliability map generation unit 221 associates a correction coefficient having a high reliability with the pixel block _BL-1 . Accordingly, the reliability map generation unit 221 generates a color calibration reliability map in which the reliability and the correction coefficient are associated with each pixel block _BL-1 . Then, the reliability map generation unit 221 generates reliability map information regarding the generated color calibration reliability map, and outputs the reliability map information to the color calibration execution unit 206 and the parallax detection unit 4.

[Configuration of color calibration execution unit]
Color calibration execution unit 206 (correction execution unit) shown in FIG. 6 acquires an input image V _L, V _R of the current frame. Further, the color calibration execution unit 206 determines which correction coefficient to use for each pixel block _BL-1 based on the reliability map information. Then, the color calibration execution unit 206 executes color calibration for each pixel block _BL-1 based on the linear correction coefficient information and the non-linear correction coefficient information.

Specifically, the color calibration execution unit 206 extracts pixels from the pixel block _BR-1 , and uses the approximate expression g ₁ (R, coef) described above to calculate the brightness of the extracted pixels and the determined correction coefficient. The value obtained by substituting is calculated as the luminance of the pixel. Color calibration execution unit 206, the processing that is performed for all the pixels constituting the input image V _R, perform color calibration. Color calibration executing section 206, the input image V _L already color _calibration, the V _R and the color calibrated image V _{L-1, V R-1,} and outputs the geometry calibration section 3 shown in FIG.

FIG. 15A shows a multi-viewpoint image V _V 1 generated by the autostereoscopic display device without executing color calibration. Similarly, FIG. 15B shows a multi-viewpoint image V _V2 generated after the autostereoscopic display device performs color calibration. In the multi-view image V _V 1, the subject image V _V 11 is broken, but in the multi-view image V _V 2, the subject image V _V 12 is not broken. Therefore, the autostereoscopic display device can generate a multi-viewpoint image without failure by executing color calibration.

[Effects of color calibration unit]
Next, effects of the color calibration unit 2 will be described. In the following description, some effects will be described in comparison with Japanese translations of PCT publication No. 2007-535829 and JP-A-2009-122842. The technique described in JP-T-2007-535829 generates a color correspondence table by generating a histogram showing a color distribution for each image and matching these histograms by DP matching. The technique described in JP-T-2007-535829 corrects the color of each image based on the color correspondence table. On the other hand, the technique described in Japanese Patent Laid-Open No. 2009-122842 extracts feature points (for example, a point on which a person's shoulder is drawn) from a plurality of images, and performs color conversion based on the color correspondence between the feature points. The model formula for is calculated. And the technique of Unexamined-Japanese-Patent No. 2009-122842 correct | amends the color of each image based on the calculated model formula.

Color calibration unit 2 on the basis of the color pair information, calculates a model equation for converting the luminance of the input image V _R to the luminance of the input image V _L. Here, the color pair information includes luminance information possessed by pixels other than the feature points. Since the color calibration unit 2 applies an IIR filter to this model formula, it is possible to calculate a robust model formula, particularly a temporally robust model formula. Therefore, the color calibration unit 2 can correct color information that is more robust than conventional ones.

  Further, the technique described in Japanese Patent Laid-Open No. 2009-122842 has a problem that the correction accuracy is low because the color conversion parameter is expressed by a simple gain or polynomial. For example, the color conversion parameter cannot express the change of the gamma value. On the other hand, the model equation of this embodiment includes at least a gain and a gamma value as parameters, so that the color calibration unit 2 can perform color calibration, that is, color information correction with high accuracy. In particular, since Equation (7) includes an offset term in addition to these as parameters, more accurate color information correction can be performed.

The color calibration unit 2 divides the input image _V L, a _{V R} to a plurality of pixel blocks _{B L-1, B R-} 1 , lowering the gradation of the pixel block _{B L-1, B R-} 1 . Then, the color calibration unit 2 generates color histograms CH _L and CH _R for each of the pixel blocks B _L−1 and B _{R −1} , and sets color pair information based on these color histograms CH _L and CH _R. Generate. Therefore, color calibration unit 2, based on the pixel block _{B L-1, B R-} 1 has been dropped tone, because it produces a color pair information (color correction information) lighter color calibration processing load Can be performed. Further, since the color calibration unit 2 matches the color histograms CH _L and CH _R and calculates the color pair information based on the result, the color pair information can be calculated with higher accuracy. In particular, since the color calibration unit 2 matches the color histograms CH _L and CH _R by DP matching, more accurate matching can be performed.

In addition, since the technique described in JP-T-2007-535829 generates a histogram indicating the color distribution of the entire image, the amount of data in the histogram increases and the processing load increases as the resolution of the image increases. There was a problem. In contrast, color calibration unit 2, so dropping the gradation of the pixel block _{B L-1, B R-} 1, the input image V _L, regardless of the resolution of V _R, processing lighter color calibration load ( Correction of color information).

The color calibration section 2 covers by generating color pair information loss of accuracy due to dropping the gradation of the pixel block _{B L-1, B R-} 1, for each pixel block B _L-1 be able to.

Furthermore, since the color calibration unit 2 performs nonlinear calculation using the parameters calculated by the linear calculation, that is, the color correction coefficients w _i and w _a as initial values, it is possible to calculate a more accurate parameter, that is, a model formula. it can.

  Furthermore, the color calibration unit 2 calculates the reliability for each of the parameter calculated by the linear calculation and the parameter calculated by the non-linear calculation, and performs color calibration based on the parameter having the higher reliability. Therefore, it is possible to correct color information with high accuracy.

<2-2. Configuration of geometric calibration unit>
Next, the configuration of the geometric calibration unit 3 will be described with reference to the drawings. As shown in FIG. 16, the geometric calibration unit 3 includes a feature amount matching unit 301, a corresponding point refinement unit 302, a linear geometric parameter fitting unit 303, a nonlinear geometric parameter fitting unit 304, and a parameter selection unit 305. And a geometric calibration execution unit 306.

[Configuration of feature matching unit]
The feature amount matching unit 301 calculates a feature amount vector by matching the luminances of the color calibrated images V _L-1 and V _R-1 . Specifically, the feature amount matching unit 301 divides the color calibrated images V _L-1 and V _R-1 into 64 pixel blocks B _L-2 and B _R-2 , respectively. Here, each of the pixel blocks B _L-2 and _BR-2 is rectangular and all have the same size. Further, the feature amount matching unit 301 extracts a pixel that is a center point of each of the pixel blocks B _L-2 and B _{R -2} , and uses the coordinates of the extracted pixels as the pixel blocks B _L-2 and B _R-2. Coordinates. Accordingly, each pixel block B _L-2 and each pixel block B _R-2 have a one-to-one correspondence. The feature amount matching unit 301 extracts feature points having a luminance equal to or higher than a predetermined value (for example, 32 to 1024) for each of the pixel blocks _BL-2 and BR _-2 . Hereinafter, the feature points of the color calibrated image V _L-1 are also referred to as “left side feature points”, and the feature points of the color calibrated image V _R-1 are also referred to as “right side feature points”. Next, the feature amount matching unit 301 performs the following processing for each left-side feature point. That is, the feature amount matching unit 301 exists in the vicinity of the left feature point (that is, the absolute value of the x coordinate difference is 64 to 256 or less and the absolute value of the y coordinate difference is 8 to 64 or less. ) Of the right feature points, the feature point having the most similar luminance (that is, the smallest absolute value of the luminance difference) is extracted, and the right feature point and the left feature point are used as a feature point pair.

Then, the feature quantity matching unit 301 calculates a feature quantity vector VE ₁ based on the feature point pair. Here, the starting point of the feature vector VE ₁ is the coordinates (u _L , v _L ) of the left feature point, and the components are (u _R −u _L , v _R −v _L ). u _R is the x coordinate of the right feature point, and v _R is the y coordinate of the right feature point. FIG. 17 shows an example of the feature vector VE ₁ and the pixel block B _L-2 . Here, the pixel block B _L-2 divides the color calibrated image V _L-1 into four.

As shown in FIG. 17, a large number of feature vectors VE ₁ having the same components exist in a portion where some object is drawn. The feature amount matching unit 301 generates feature amount vector information regarding the pixel blocks B _L-2 and _BR-2 and the feature amount vector VE ₁ and outputs the feature amount vector information to the corresponding point refinement unit 302 and the nonlinear geometric parameter fitting unit 304. In the present embodiment, since geometric calibration is performed after color calibration, the feature amount matching unit 301 can perform luminance matching with high accuracy. For this reason, the accuracy of geometric calibration is improved. Of course, the geometric calibration may be performed before the color calibration. In this case, it is possible to perform luminance matching with high accuracy in color calibration, and as a result, the accuracy of color calibration is improved.

[Configuration of corresponding point refinement section]
As shown in FIG. 18, the corresponding point refinement unit 302 includes a histogram generation unit 307, a filter unit 308, a maximum frequency detection unit 309, a coordinate assignment unit 310, and a corresponding point selection unit 311.

The histogram generation unit 307 generates a vector component histogram indicating the frequency of each component of the feature vector VE ₁ for each pixel block B _L-2 based on the feature vector information. FIG. 19 shows an example of a vector component histogram. x-axis, y-axis is the x component of each feature vector VE _1, it shows the y component, z-axis indicates the frequency of each component. When any object is drawn in the pixel block B _L-2 , a large number of feature amount vectors VE ₁ having the same components exist in the portion where the object is drawn. Therefore, in this case, the vector component histogram has a large peak. The histogram generation unit 307 generates vector component histogram information related to the vector component histogram and outputs the vector component histogram information to the filter unit 308.

  The filter unit 308 smoothes the vector component histogram by applying a Gaussian filter to the vector component histogram information a plurality of times. The filter unit 308 generates smoothed histogram information related to the smoothed vector component histogram, and outputs the smoothed histogram information to the maximum frequency detection unit 309 and the corresponding point selection unit 311.

The maximum frequency detection unit 309 specifies an x component and a y component with the maximum frequency from each vector component histogram based on the smoothed histogram information. The feature vector VE ₁ having such x component and y component is set as the representative vector VE _{2 of the} pixel block _BL-2 to which the feature vector VE ₁ belongs. Thus, the maximum frequency detecting unit 309 calculates a representative vector VE ₂ for each pixel block _{B L-2.} The maximum frequency detection unit 309 generates representative vector component information regarding the component of the representative vector VE ₂ and outputs the representative vector component information to the coordinate assigning unit 310.

Coordinate giving unit 310, based on the representative vector component information to determine the coordinates each representative vector VE ₂ (start point coordinates). That is, the representative vector VE ₂ is calculated from the vector component histogram. However, since the vector component histogram does not have the coordinate information of each feature quantity vector VE ₁ , the maximum frequency detection unit 309 performs the representative vector VE _2. at the time of calculating the representatively vector VE ₂ has no coordinate information. Therefore, the coordinate giving unit 310, based on the representative vector information to determine the coordinates each representative vector VE ₂ (start point coordinates). Specifically, as shown in FIG. 20, the coordinate assigning unit 310 sets the center point P _L-1 11 of the pixel block B _L-2 as the initial value of the coordinates of the representative vector VE ₂ and repeats the mean shift process. . Here, Mean shift processing includes the coordinates representative vector VE _2, calculates the average value of the coordinates feature vector VE ₁ having the same components as the representative vector VE _2, new representative vector VE ₂ and the average value It is a process to make a coordinate. Therefore, by Mean shift processing is repeated, the coordinates representative vectors VE _2, the area where the feature quantity vector VE ₁ having the same components as the representative vector VE ₂ is dense, i.e. moves in the dense region. Coordinate giving unit 310, the coordinate representative vectors VE ₂ performs repeatedly Mean shift processing until convergence (e.g., about 10 times). Thus, the coordinate assigning unit 310 determines the coordinates representative vector VE _2. At this time, the coordinate representative vectors VE ₂ has a center _{P L-1} 21 dense region. Then, the coordinate assigning unit 310 generates representative vector information regarding the components and coordinates of the representative vector VE ₂ and outputs the representative vector information to the corresponding point selection unit 311.

The corresponding point selection unit 311 deletes a plurality of representative vectors VE ₂ whose maximum frequency is less than a predetermined value (for example, 0.1 to 0.4) based on the representative vector information and the smoothed histogram information. Do (ie, cut off). Thereby, the corresponding point selection unit 311 can exclude the representative vector VE ₂ having low reliability, for example, the representative vector VE ₂ calculated from the pixel block _{BL-2 on} which no object is drawn. Then, the corresponding point selection unit 311 generates corresponding point selection information related to the feature point pair constituting the selected representative vector VE ₂ , that is, the representative point pair, and outputs it to the linear geometric parameter fitting unit 303. Hereinafter, among the coordinate points constituting the representative point pair, the left coordinate point is also referred to as the left representative point, and the right coordinate point is also referred to as the right representative point. The coordinates of the left representative point are the starting points (u _L , v _L ) of the representative vector, and the coordinates of the right representative point are (u _R , v _R ) obtained by adding the coordinates of the starting point to the representative vector components.

The reason why the representative vector VE ₂ is extracted from the feature vector VE ₁ is that the linear operation is weak to the outlier. That is, as will be described later, the linear geometric parameter fitting portion 303, by performing a linear operation using a representative vector VE _2, calculates a transformation matrix F. This transformation matrix F is used for coordinate transformation of the color calibrated images V _L-1 and V _R-1 . On the other hand, since the feature vector VE ₁ is simply obtained by connecting the feature points, the value (component) is likely to vary (that is, likely to vary). Therefore, if a linear operation using the feature vector VE ₁ is performed, each component of the transformation matrix F obtained by the linear operation is likely to be exposed. Therefore, in this embodiment, the representative vector VE ₂ is extracted from the feature vector VE ₁ .

[Configuration of linear geometric parameter fitting unit]
The linear geometric parameter fitting unit 303 (conversion matrix calculation unit) includes an initialization unit 312, a weight calculation unit 313, and a linear geometric parameter calculation unit 314.

  First, an outline of processing performed by the linear geometric parameter fitting unit 303 will be described. The linear geometric parameter fitting unit 303 calculates the transformation matrix F that satisfies the epipolar constraint condition and the time axis constraint condition, and decomposes the transformation matrix F to calculate the homography matrices HL and HR.

Here, the transformation matrix F is a 3 × 3 matrix and includes homography matrices HL and HR. The homography matrix HL is a matrix for performing the coordinate transformation of the color calibrated image V _L-1 , and the homography matrix HR is a matrix for performing the coordinate transformation of the color calibrated image VR _−1. is there.

The epipolar constraint condition is a constraint that the left coordinate point is close to (preferably rides on) the epipolar line obtained by converting the right coordinate point by the transformation matrix F, in other words, the vector u ′ (u _R , v _R , 1). This is a constraint that the inner product of the epipolar vector obtained by multiplying the matrix F and the vector u (u _L , v _L , 1) becomes zero. Here, (u _L , v _L ) are the coordinates of the left coordinate point, and (u _R , v _R ) are the coordinates of the right coordinate point.

  The epipolar constraint condition is expressed by the following equation (19).

Here, f _{0 to} f ₈ are each component of the transformation matrix F. The linear geometric parameter fitting unit 303 converts the epipolar constraint condition into a least square problem expressed by the following equation (20) by tearing the equation (19).

Here, M is an operation matrix and is represented by the following equation (21).

The coordinates of different representative point pairs are substituted into (u _L , v _L ) and (u _R , v _R ) of each row of the operation matrix M.

On the other hand, the time axis constraint condition is that the current frame conversion matrix F _n is constrained to the previous frame conversion matrix F _n−1 , that is, the current frame conversion matrix F _n and the previous frame conversion matrix F _n. The constraint is that the correlation with _-1 is maximized. Since the transformation matrix F is a unit vector, these inner products are directly correlated.

  Since the transformation matrix F calculated without using such a time axis constraint condition often fluctuates in the time axis direction (the value varies depending on each time point), in this embodiment, the time axis constraint condition is By using this to calculate the transformation matrix F, the fluctuation of the transformation matrix F in the time axis direction is reduced. This improves the accuracy of the transformation matrix F. The time axis constraint condition is expressed by the following equation (22).

  The time axis constraint condition equation (22) is replaced with a maximum square problem represented by the following equation (23).

  Therefore, the linear geometric parameter fitting unit 303 solves the mixed problem of the least square problem expressed by the equation (20) and the maximum square problem expressed by the equation (23). Cannot be solved by linear calculation.

On the other hand, the fact that the inner product of the transformation matrix _{F n} and transformation matrix _{F n-1} is 1, the inner product is 0 and the orthogonal vector group F _'n-1 of the transformation matrix _{F n} transformation matrix _{F n-1} It is synonymous with becoming. Therefore, the maximum square problem of Expression (23) is replaced with the least square problem represented by the following Expression (24).

  Therefore, the epipolar constraint condition and the time axis constraint condition can be summarized as a least square problem expressed by the following equation (25).

However, since the transformation matrix F _n−1 is a random value, it is not easy to directly calculate the orthogonal vector group F ′ _n−1 of the transformation matrix F _n−1 .

  On the other hand, the transformation matrix F is expressed as follows using the homography matrices HL and HR.

Here, the matrix F ₀ is a vector composed of only fixed values, and is represented by the following equation (27).

In the equation (19), when the homography matrices HL and HR are respectively multiplied by the three-dimensional coordinates (u _L , v _L , 1) and (u _R , v _R , 1) of the representative point pair, the equation (19) becomes Is converted as follows.

Here, (u ′ _L , v ′ _L , 1), (u ′ _R , v ′ _R , 1) are the three-dimensional coordinates (u _L , v _L , 1), (u _R , v) of the representative point pair. _R (1) is a coordinate (conversion vector) obtained by multiplying the homography matrix HL, HR by the homography matrix HL, HR (that is, coordinate transformation by the homography matrix HL, HR). When equation (25) is rewritten using equation (28), it becomes as follows.

  Here, M ′ is obtained by changing each component of the arithmetic matrix M described above to a value after coordinate conversion, and is specifically expressed by the following equation (30).

The matrix F ′ ₀ is an orthogonal vector group of the matrix F ₀ . Since the matrix F ₀ is a nine-component matrix, there are eight orthogonal vectors of the matrix F ₀ . Since the matrix F ₀ is a fixed value, the orthogonal vector group is also a fixed value. Therefore, the linear geometric parameter fitting unit 303 does not need to calculate the matrix F ′ ₀ . Specifically, the linear geometric parameter fitting unit 303 stores a matrix F ′ ₀ in advance.

Also, the linear geometric parameter fitting portion 303, instead of directly using the time axis constraint, used multiplied by the weight w _e to the time axis constraint. This weight _{w e,} equation (29) can be rewritten as follows.

Therefore, the linear geometric parameter fitting unit 303 calculates the matrix F ′ _n (fixed value equivalent vector) by solving the least square problem of Equation (31), and thus calculates the transformation matrix F. The linear geometric parameter fitting unit 303 calculates the homography matrices HL and HR by decomposing the transformation matrix F. By converting the left coordinate point of the color calibrated image V _{L-1 with} the homography matrix HL and converting the right coordinate point of the color calibrated image VR _{-1 with} the homography matrix HR, Is reduced. Thus, in this embodiment, constraint that is to calculate the epipolar constraint and the time axis satisfying the constraint condition transformation matrix F, other constraints, for example, an input image V _L, V _R are parallel to each other May be added as a constraint condition. Hereinafter, each configuration of the linear geometric parameter fitting unit 303 will be described in detail.

[Configuration of initialization section]
As illustrated in FIG. 22, the initialization unit 312 includes an error calculation unit 315, an initial parameter selection unit 316, a coordinate pair holding unit 317, and a coordinate pair integration unit 318.

The error calculation unit 315 performs error calculation based on the corresponding point selection information and parameter information given from the parameter selection unit 305 described later. Here, the parameter information includes information on the homography matrices HL _n−1 and HR _n−1 calculated in the previous frame.

  Specifically, the error calculation unit 315 calculates error values e1 and e2 by the following equations (32) and (33).

Here, u and u ′ are three-dimensional representations of the coordinates of the representative point pairs, that is, (u _L , v _L , 1), (u _R , v _R , 1), and I is a unit matrix. Σ means summing values for all representative point pairs.

Equations (32) and (33) are the same as the equation (19) of the epipolar constraint, but the homography matrices HL and HR of the transformation matrix F are the homography matrices HL _n−1 and HR _{n−1 of the} previous frame, respectively, or the unit matrix It corresponds to the one replaced with I.

Therefore, the smaller the error values e1 and e2, the better the accuracy of the homography matrices HL _n−1 and HR _n−1 and the unit matrix I as the homography matrices HL and HR.

  Generally, when the same scene is continuous, the error value e1 becomes smaller than the error value e2, but when a scene change occurs, the error value e2 becomes smaller than the error value e1.

  Then, the error calculation unit 315 generates error value information regarding the calculated error values e1 and e2, and outputs the error value information to the initial parameter selection unit 316 together with the corresponding point selection information. In addition, the error calculation unit 315 selects an error value having a small value, and generates minimum error value information regarding the selected error value. Then, the error calculation unit 315 outputs the minimum error value information and the corresponding point selection information to the weight calculation unit 313.

Based on the error value information, the initial parameter selection unit 316 determines which one of the previous frame homography matrices HL _n−1 , HR _n−1, and the unit matrix I is used as an initial parameter. Specifically, the initial parameter selection unit 316 sets the matrix with the smaller error value as the initial parameter homography matrix.

  The initial parameter selection unit 316 generates initial parameter information related to the selected initial parameter and outputs it to the linear geometric parameter calculation unit 314. Further, the initial parameter selection unit 316 outputs the corresponding point selection information to the coordinate pair holding unit 317 and the coordinate pair integration unit 318. The initial parameter selection unit 316 outputs reset information to the coordinate pair holding unit 317 when a unit matrix is selected as the initial parameter.

  The coordinate pair holding unit 317 holds the corresponding point selection information given from the initial parameter selection unit 316, and out of the held corresponding point selection information, the corresponding point selection information for the past four frames is sent to the coordinate pair integration unit 318. Output. The coordinate pair holding unit 317 deletes the corresponding point selection information when the reset information is given from the initial parameter selection unit 316. This is because when the reset information is given from the initial parameter selection unit 316, it is highly possible that there has been a scene change, so it is preferable that the corresponding point selection information of the past frame is not used.

  The coordinate pair integration unit 318 integrates the corresponding point selection information given from the initial parameter selection unit 316 and the coordinate pair holding unit 317. Thereby, the coordinate pair integration unit 318 integrates corresponding point selection information for five frames at the maximum. The coordinate pair integration unit 318 outputs the integrated corresponding point selection information to the weight calculation unit 313 and the linear geometric parameter calculation unit 314.

[Configuration of Weight Calculation Unit]
Based on the minimum error value information and the corresponding point selection information given from the error calculation unit 315, the weight calculation unit 313 indicates a weight indicating how much the time axis constraint condition is considered when calculating the transformation matrix F, that is, to calculate the weight w _e of the time axis constraint.

Specifically, the weight calculation unit 313 includes an error value (initial value error in a frame), the number of representative vectors VE ₂ (number of samples), and a range in which the variance of the representative vector VE ₂ is 0.0 to 3.0 ( when the first range) within determines the weight w _e time axis constraint to the first weight w _e1, the error value, the number of representative vectors VE _2, and dispersion of the representative vector VE ₂ 3 range .0～16.0 when the (second range) within the weight _{w e} was determined second weight _{w e,} error values, the number of representative vectors VE _2, and representative vectors VE ₂ If the dispersion is in the range of 16.0～ (third range) determines the weight _{w e} to the third weight _{w e3.} The magnitude relationship of each _weight, the _{w e1> w e3> w e2} . The value of the first weight _{w e1} is from 10.0 to 1.0 for example, the value of the second weight _{w e2} is from 1.0 to 0.1 for example, the value of the third weighting _{w e3,} for example 0.1 to 0.01.

The reason why this determination is possible will be described with reference to FIG. Figure 23 is a graph showing error values, the number of representative vectors VE _2, and a dispersion of the representative vector VE _2, and the weight of the time axis constraint, the relationship between the accuracy of the transformation matrix F is calculated. That is, each point in the graph shows the error values, the number of representative vectors VE _2, and the dispersion of the representative vector VE ₂ relationship. These points are indicated by a point P _w1 when the weight is higher (for example, 1) and the accuracy of the transformation matrix F is better, and the weight is lower (for example, 0.01) than the transformation matrix F. When the accuracy is good, it is indicated by a point _Pw2 .

According to FIG 23, an error value representative number of vectors VE _2, and by the dispersion of the representative vector VE _2, have different weights suitable for the calculation of the transformation matrix F. Further, the point P _w1 is distributed in the areas AR1 and AR2, and the point P _w4 is distributed in the area AR3. Note that the point P _w1 is also distributed in the region AR4 that is distant from the regions AR1 and AR2, but this is due to lens distortion (blow-fish).

Areas AR1 and AR2 indicate the first range described above, and area AR3 indicates the second range described above. Therefore, the weight calculation section 313, an error value, the number of representative vectors VE _2, and representative if the variance of the vector VE ₂ belongs to the area AR1 or area AR2, i.e., when a in the first range, the time axis constraint determining a weight _{w e} conditions the heaviest first weight _{w e1.} The weight calculation section 313, an error value, if the number of representative vectors VE _2, and dispersion of the representative vector VE ₂ belongs to the area AR3, i.e. if the second range is, weighting the time axis constraint to determine the w _e to the lightest second weight _{w e2.} For example, a point included in the area AR1 indicates that the representative point of the current frame is geometrically corrected with a small error value, that is, the homography matrices HL _n−1 and HR _{n−1 of the} previous frame. Thus, an error value, if the number of representative vectors VE _2, and dispersion of the representative vector VE ₂ belongs to the area AR1, and the transformation matrix _{F n} of the current frame is close to the transformation matrix _{F n-1} of the previous frame (i.e., It can be said that the transformation matrix F has converged). On the other hand, the points included in the area AR3 indicate that even if the feature points of the current frame are geometrically corrected with a large error value, that is, the homography matrices HL _n-1 and HR _{n-1 of the} previous frame, the accuracy is poor. Therefore, the error value, the number of representative vectors VE _2, and if the variance of the representative vector VE ₂ belongs to the area AR3, no transformation matrix _{F n} of the current frame is close to the transformation matrix _{F n-1} of the previous frame (i.e., It can be said that the conversion matrix F has not converged).

On the other hand, the error value, the number of representative vectors VE _2, and if the variance of the representative vector VE ₂ belongs to the area AR4 is the lens is distorted, the weight w _e time axis constraint first weight w _It may be desirable to make it different from _e1 and the second weight w _e2 . Therefore, the weight calculation section 313, an error value, if the number of representative vectors VE _2, and dispersion of the representative vector VE ₂ belongs to the area AR4, i.e. if the third range is weighted in the time axis constraint determining w _e to the third weight _{w e3} and first weight _{w e1} is intermediate between the second weight _{w e2.}

Weight calculator 313 generates a time axis constraint weight information about the weight w _e of determined time axis constraint, and outputs the linear geometric parameter calculation unit 314.

[Configuration of linear geometric parameter calculation unit]
As shown in FIG. 24, the linear geometric parameter calculation unit 314 includes a coordinate conversion unit 319, a matrix calculation unit 320, an F matrix calculation unit 321, and an F matrix decomposition unit 322.

The coordinate conversion unit 319 performs coordinate conversion of each representative point pair based on the information given from the initialization unit 315. Specifically, the coordinate conversion unit 319 performs coordinate conversion of the left representative point with the initial parameter, that is, the homography matrix HL _n-1 or the unit matrix I of the previous frame, and converts the right representative point to the initial parameter, that is, the previous frame homology. Coordinate conversion is performed using the graphy matrix HR _n-1 or the unit matrix I. In the unit matrix I, the coordinate conversion is not substantially performed.

Such coordinate conversion may cause the center points and coordinate ranges of the color calibrated images V _L-1 and V _R-1 to deviate. Therefore, the coordinate conversion unit 319 normalizes the coordinates and initial parameters of the representative point pair after the coordinate conversion. Specifically, for example, the coordinate conversion unit 319 multiplies the three-dimensional coordinates and initial parameters of each representative point pair by a conversion matrix for normalization. Hereinafter, the normalized three-dimensional coordinates are also referred to as normalized coordinates, and the normalized initial parameters are also referred to as normalized homography matrices WL and WR.

  The coordinate conversion unit 319 generates coordinate conversion information regarding the normalized coordinates of each representative point pair and the normalized homography matrices WL and WR, and outputs the coordinate conversion information to the matrix calculation unit 320.

Matrix calculation unit 320, based on the coordinate transformation information, and calculates the operation matrix M ₂ represented by the following formula (34).

Here, Σ means summing up for all representative point pairs, and the matrix C is a combination of the above-described orthogonal vector group F ′ ₀ and is expressed by the following equation (35).

The matrix calculation unit 320 generates calculation matrix information regarding the calculated calculation matrix M ₂ and outputs the calculation matrix information to the F matrix calculation unit 321.

The F matrix calculation unit 321 calculates the minimum eigenvector v _min of the operation matrix M ₂ based on the operation matrix information. The method for calculating the minimum eigenvector v _min is not particularly limited, but singular value decomposition or the like is preferable, for example.

Further, F matrix calculation unit 321, so that the minimum singular value of the minimum eigenvector _{v min} drops to 0, singular value decomposition minimum eigenvector _{v min} in a matrix of 3 × 3. Then, the F matrix calculation unit 321 calculates the transformation matrix F based on the calculated matrix SVD (v _min ) and the following equation (36).

Further, F matrix calculation unit 321 by shaking the weight _w value several values of _e, for example, _{^{2 2 × w e, 2 ×}} w e, 2 -1 × w e, 2 -2 × w e, converted The matrix F is calculated. Then, the F matrix calculation unit 321 calculates an error value for each of the calculated conversion matrices F using the above-described equation (32). Then, the F matrix calculation unit 321 selects the one with the smallest error value, generates transformation matrix information regarding the selected transformation matrix F, and outputs the transformation matrix information to the F matrix decomposition unit 322 together with the coordinate transformation information.

The F matrix decomposition unit 322 decomposes the transformation matrix F into homography matrices HL, HR, and a matrix F ₀ based on the transformation matrix information. Substantially, the F matrix decomposition unit 322 calculates matrices that make the epipolar lines parallel as the homography matrices HL and HR.

Specifically, the F matrix decomposition unit 322 first calculates the epipole point coordinates by performing epipole calculation. An epipole point is an intersection of epipolar lines. Assuming that the light rays incident on the image pickup apparatus are collected at one point, straight lines obtained by projecting these light rays onto the color calibrated images V _L-1 and V _R-1 , that is, epipolar lines are also one epipole point. Will be exchanged. Since the epipole point is on an arbitrary epipolar line, the inner product for the arbitrary epipolar line becomes zero.

Therefore, the F matrix decomposition unit 322 calculates the coordinates of the epipole points such that the square of the inner product of the epipolar line and the coordinates of the epipole points is minimized. Specifically, the F matrix decomposition unit 322 calculates the square of the transformation matrix F. Next, the F matrix decomposition unit 322 sets the calculated minimum eigenvector of the matrix as the coordinates of the epipole point. Since the epipole points exist in each of the color calibrated images V _L-1 and V _R-1 , the coordinates of these epipole points are calculated. That is, the F matrix decomposition unit 322 calculates the minimum eigenvector el of the matrix FL represented by the following equation (37), and uses this to calculate the epipole point on the color calibrated image V _L−1 , that is, the left epipole point. Use coordinates.

  Here, t is transposition.

Similarly, the F-matrix decomposition unit 322 calculates the minimum eigenvector eR of the matrix FR represented by the following equation (38), and uses this as the epipole point on the color calibrated image VR _-1 , that is, the right epipole point. Coordinates.

Then, the F matrix decomposition unit 322 calculates a matrix that skips the left epipole point to infinity (that is, makes the epipolar line on the color calibrated image V _L-1 parallel) as the homography matrix HL, and the right epipole A matrix that skips points to infinity (that is, makes the epipolar line on the color calibrated image VR _-1 parallel) is calculated as a homography matrix HR.

  Specifically, the F matrix decomposition unit 322 calculates the projection component matrix Hp of the homography matrices HL and HR by the following equations (39) and (40).

  Here, e (2) is the x coordinate of the epipole, and e (0) is the z coordinate of the epipole.

  Then, the F matrix decomposition unit 322 performs Rectification. The rotation component matrix Hr of the homography matrix HL and HR is calculated using calcRotationOfRectify of cpp, and Rectification. The distortion (translation) component matrix Hsh of the homography matrices HL and HR is calculated by using calc sharingOfRectify of cpp. Then, the F matrix decomposition unit 322 calculates the vertical scale component matrix Hs of the homography matrices HL and HR based on the following equations (41) and (42).

  Here, “Vertical Torrance” is a horizontal translational movement amount, “Average Error” is a difference average with respect to the movement amount, and σ is a standard deviation.

  Then, the F matrix decomposition unit 322 calculates the homography matrix HL by multiplying the projection component matrix Hp, the rotation component matrix Hr, the distortion component matrix Hsh, and the vertical scale component matrix Hs of the homography matrix HL.

  Similarly, the F matrix decomposition unit 322 calculates the homography matrix HR by multiplying the projection component matrix Hp, the rotation component matrix Hr, the distortion component matrix Hsh, and the vertical scale component matrix Hs of the homography matrix HR.

  Next, the F matrix decomposition unit 322 generates linear correction coefficient information regarding the calculated homography matrices HL and HR, and outputs the linear correction coefficient information to the nonlinear geometric parameter fitting unit 304 and the parameter selection unit 305.

[Configuration of nonlinear geometric parameter fitting unit]
As shown in FIG. 25, the nonlinear geometric parameter fitting unit 304 includes a Jacobian matrix calculating unit 323, a Hessian matrix calculating unit 324, and a coefficient updating unit 325.

The Jacobian matrix calculation unit 323 is based on the feature quantity vector information given from the feature quantity matching unit 301 and the linear correction coefficient information given from the linear geometric parameter fitting unit 303, and is represented by the following equation (43). to calculate the matrix _{J 2.}

Here, u is a vector indicating the three-dimensional coordinates of the left feature point, u ′ is a vector indicating the three-dimensional coordinates of the right feature point, and coef (F) indicates each component of the transformation matrix F. f _e2 (u, u ′, coef (F)) is a so-called error function, and is represented by the following equation (44). f _{e20 to} f _e2m are obtained by substituting different feature vector information into the error function f _e2 (u, u ′, coef (F)).

Here, g ₂ (u, coef (F)) is a value obtained by multiplying the three-dimensional coordinates of the left feature point by the transformation matrix F, and g ₂ (u ′, coef (F)) is the right feature point. It is a value obtained by multiplying the three-dimensional coordinates by the transformation matrix F.

  RobustFunc is a function represented by the above-described equation (16).

Then, the Jacobian matrix calculation unit 323 generates Jacobian matrix information regarding the calculated Jacobian matrix J ₂ and outputs the generated Jacobian matrix information to the Hessian matrix calculation unit 324.

Further, when the Jacobian matrix calculation unit 323 is provided with nonlinear correction coefficient information from a coefficient update unit 325 described later, the Jacobian matrix calculation unit 323 includes a nonlinear correction coefficient coef (nonlinear) indicated by the nonlinear correction coefficient information and feature quantity vector information. based on, for calculating the Jacobian matrix J ₂ described above. Then, the Jacobian matrix calculation unit 323 generates Jacobian matrix information regarding the calculated Jacobian matrix J ₂ and outputs the generated Jacobian matrix information to the Hessian matrix calculation unit 324.

Hessian matrix calculation section 324, based on the information provided by the Jacobian matrix calculation section 323 calculates the Hessian matrix H ₂ represented by the following formula (45).

Then, Hessian matrix computation unit 324 generates the Hessian information about Hessian H ₂ calculated, and outputs the coefficient update unit 325 together with the information provided by the Jacobian matrix calculation section 323.

  The coefficient updating unit 325 calculates each component of the transformation matrix F as a nonlinear correction coefficient coef (nonlinear) based on the information given from the Hessian matrix calculation unit 324 and the following equation (46).

Here, I is a unit matrix. w _f is a coefficient (weight), for example, 0.1.

  Next, the coefficient updating unit 325 calculates the homography matrices HL and HR by decomposing the nonlinear correction coefficient coef (nonlinear) by the same process as the process of the F matrix decomposing unit 322 described above. Then, the coefficient updating unit 325 generates nonlinear correction coefficient information regarding the calculated homography matrices HL and HR. Then, the coefficient updating unit 325 determines whether the homography matrices HL and HR have converged to a constant value. If the homography matrices HL and HR have converged to a certain value, the nonlinear correction coefficient information is selected as a parameter. When it is determined that the homography matrices HL and HR have not converged, the nonlinear correction coefficient information is output to the Jacobian matrix calculation unit 323. Thereafter, the homography matrices HL and HR are recalculated. Therefore, the non-linear geometric parameter fitting unit 304 calculates the homography matrices HL and HR by a non-linear calculation method using the homography matrices HL and HR calculated by the linear geometric parameter fitting unit 303 as initial values, and obtains the homography matrix HL, The nonlinear calculation is repeated until HR converges to a constant value.

[Configuration of parameter selection section]
The parameter selection unit 305 illustrated in FIG. 16 uses the above-described equation (32) for each of the linear correction coefficient information given from the linear geometric parameter fitting unit 303 and the nonlinear correction coefficient information given from the nonlinear geometric parameter fitting unit 304. ) To calculate an error value for each set of pixel blocks B _L−2 and B _R−2 . Then, the parameter selection unit 305 selects information having a smaller error value as parameter information. The parameter selection unit 305 selects parameter information for each set of pixel blocks B _L−2 and B _R−2 and outputs the parameter information to the geometric calibration execution unit 306 and the linear geometric parameter fitting unit 303.

[Configuration of Geometric Calibration Execution Unit 306]
The geometric calibration execution unit 306 performs geometric calibration on the color calibrated images V _L-1 and V _R-1 based on the parameter information. Specifically, the geometric calibration execution unit 306 performs coordinate conversion of each coordinate point of the color calibrated image V _L-1 using the homography matrix HL, and converts each coordinate point of the color calibrated image V _R-1. Coordinate transformation is performed using the homography matrix HR. The geometric calibration execution unit 306 outputs the images after geometric calibration to the parallax detection unit 4 shown in FIG. 5 as the calibrated images V _L-2 and V _R-2 , respectively.

  FIG. 26 shows how the variation of the transformation matrix F is reduced by using the time axis constraint condition. A curve L3 in FIG. 26A shows how the components of the transformation matrix F calculated without using the time axis constraint condition change, and a curve L4 in FIG. 26B uses the time axis constraint condition. A state in which the calculated component of the transformation matrix F varies is shown. As shown in FIG. 26, the variation of the value of the transformation matrix F calculated using the time axis constraint condition is greatly reduced.

  FIG. 27 shows how the transformation matrix F converges over time. Curves L5 and L6 show how the error value of Equation (32) converges over time. A curve L5 shows a case where the transformation matrix F is calculated using the corresponding point selection information for one frame, and a curve L6 shows the transformation matrix F calculated using the corresponding point selection information for five frames. Show the case. A curve L7 indicates an error value of Expression (33). As shown in FIG. 27, the conversion matrix F largely converges with time. Further, the transformation matrix F calculated using the corresponding point selection information for five frames converges more greatly than the transformation matrix F calculated using the corresponding point selection information for one frame, and the error value is stable. It can be said that.

[Effects of the geometric calibration unit 3]
Next, effects of the geometric calibration unit 3 will be described.

  Since the geometric calibration unit 3 calculates the transformation matrix F and performs geometric calibration (geometric correction) based on the transformation matrix F, the geometric calibration can be performed without a known pattern (known model). . Further, since the transformation matrix F satisfies the epipolar constraint condition, it can be applied to various images than the known pattern. Since the geometric calibration unit 3 solves the epipolar constraint condition using the above-described equation (19), the transformation matrix F can be calculated with high accuracy.

  In addition, the geometric calibration unit 3 calculates a transformation matrix F that satisfies not only the epipolar constraint condition but also the time axis constraint condition, so that the transformation matrix is highly accurate and robust in the time axis direction (high temporal robustness). F can be calculated. For this reason, the geometric calibration unit 3 can perform a geometric calibration that is robust in time. In addition, the present inventor has used nonlinear estimation methods such as RANSAC estimation and M estimation as geometric calibration methods (for details of these estimation methods, see, for example, “Mahamud and Hebert 2000,” Interactive projective restructuring from multiple view. ”” And “Hartley and Zisserman 2000,“ Multiple view geometry in computer vision ”” etc.) were also examined. As a result, in the RANSAC estimation, it has been found that the parameters of the transformation matrix F are likely to fluctuate (fluctuate easily) in the time axis direction. Further, it has been found that the processing load on the image processing apparatus becomes very heavy in the M estimation. On the other hand, the transformation matrix F calculated in the present embodiment is less likely to be displaced in the time axis direction, and the processing load is lighter.

  In addition, since the geometric calibration unit 3 uses the time axis constraint as the correlation with the transformation matrix F calculated in the previous frame, the transformation matrix F can be calculated with high accuracy, and thus the geometric calibration is performed with high accuracy. be able to.

Further, the geometric calibration unit 3 performs the maximum square problem of maximizing the correlation with the transformation matrix F _n−1 calculated in the previous frame, and uses the orthogonal vector group of the transformation matrix F _n−1 calculated in the previous frame. Since it is converted into a least square problem that minimizes the correlation with F ′ _n−1 , the epipolar constraint condition and the time axis constraint can be solved as a single least square problem.

Also, the geometric calibration unit 3, weighting the time axis constraint, the weight w _e time axis constraint, since the change according to the condition of the representative point pair (the number of samples, etc.), to accurately calculate a transformation matrix F As a result, the convergence of the transformation matrix F can be accelerated.

Also, the geometric calibration unit 3, swing some weight w _e time axis constraint, calculates a transformation matrix F using the respective weight w _e, since selecting a transformation matrix F where the error value is minimized, The conversion matrix F can be calculated with high accuracy.

  In addition, the geometric calibration unit 3 uses a feature point pair having a particularly high frequency as a representative point pair among the feature point pairs, and calculates the transformation matrix F based on the representative point pair. can do.

Also, the geometric calibration unit 3, the coordinate representative vectors VE ₂ so calculated by Mean shift processing, it is possible to accurately calculate the coordinates representative vector VE _2.

In addition, the geometric calibration unit 3 converts the coordinates of the left representative point and the right coordinate point with initial parameters, for example, the homography matrices HL _n−1 and HR _n−1 , before calculating the conversion matrix F. The conversion matrix F can be easily calculated. Specifically, the geometric calibration unit 3 can convert the orthogonal vector group F ′ _n−1 to a fixed value.

  Further, since the geometric calibration unit 3 performs the nonlinear calculation using the parameters of the conversion matrix F calculated by the linear calculation as initial values, the conversion matrix F can be calculated with high accuracy.

  Further, the geometric calibration unit 3 calculates an error value for each of the conversion matrix F calculated by the linear calculation and the conversion matrix F calculated by the nonlinear calculation, and uses the conversion matrix F having the smaller error value. Since the geometric calibration is performed, the geometric calibration can be performed with high accuracy.

[First Modification of Linear Geometric Parameter Fitting Unit]
Next, a first modification of the linear geometric parameter fitting unit will be described. As shown in FIG. 28, the linear geometric parameter fitting unit 326 includes an initialization unit 327, a weight calculation unit 328, a coordinate conversion unit 329, a projection component calculation unit 330, a rotation component calculation unit 331, and a translation component calculation. A unit 332 and an F matrix decomposition unit 333 are provided. In general, the linear geometric parameter fitting unit 326 calculates the transformation matrix F for each component, that is, for each projection component, rotation component, and translation component, and when calculating each component, The point which considers a component differs from the linear geometric parameter fitting part 303 mentioned above.

  The initialization unit 327, the weight calculation unit 328, and the coordinate conversion unit 329 perform the same processing as the initialization unit 312, the weight calculation unit 313, and the coordinate conversion unit 319 described above. However, the weight calculation unit 328 outputs the above time axis constraint weight information to the projection component calculation unit 330, the rotation component calculation unit 331, and the translation component calculation unit 332.

As shown in FIG. 29, the projection component calculation unit 330 includes a matrix calculation unit 334 and an F matrix calculation unit 335. The matrix calculation unit 334 performs the same process as the matrix calculation unit 320 described above. However, the matrix calculation unit 334 calculates a calculation matrix M ₃ represented by the following equation (47).

Here, Σ means summing up for all representative point pairs, and the matrix C ₁ is a combination of a part of the orthogonal vector group F ′ ₀ described above, and is expressed by the following equation (48). .

The matrix calculation unit 334 generates calculation matrix information related to the calculated calculation matrix M ₃ and outputs the calculation matrix information to the F matrix calculation unit 335.

The F matrix calculation unit 335 performs the same processing as the F matrix calculation unit 321 described above based on the calculation matrix information. Thus, to calculate a transformation matrix F _p of projection component represented by the following formula (49).

Here, Pj1 to Pj5 indicate arbitrary numerical values. The F matrix calculation unit 335 generates projection component conversion matrix information regarding the projection component conversion matrix F _p and outputs the projection component conversion matrix information to the rotation component calculation unit 331 together with the coordinate conversion information.

As illustrated in FIG. 30, the rotation component calculation unit 331 includes a matrix calculation unit 336 and an F matrix calculation unit 336. The matrix calculation unit 336 performs the same process as the matrix calculation unit 320 described above. However, the matrix calculating unit 336 calculates the operation matrix _{M 4} represented by the following formula (50).

Here, Σ means summing up for all representative point pairs, and the matrix C ₂ is a combination of a part of the above-described orthogonal vector group F ′ ₀ and is expressed by the following equation (51). .

The matrix calculation unit 336 generates calculation matrix information related to the calculated calculation matrix M ₄ and outputs the calculation matrix information to the F matrix calculation unit 337. As shown in equation (50), the operation matrix M _4, contains transformation matrix F _P of the projection component. That is, operation matrix M ₄ is adapted to that in consideration of the matrix F _P of the projection component.

The F matrix calculation unit 337 performs the same processing as the F matrix calculation unit 321 described above based on the calculation matrix information. Thus, to calculate a transformation matrix F _R of the rotation component of the following formula (52). Here, the operation matrix M _4, because it contains the transformation matrix F _P of the projection components, F matrix calculation unit 337 to calculate a transformation matrix F _R which is bound to the transformation matrix F _P of the projection component Can do.

Here, Rot1 to Rot4 indicate arbitrary numerical values. F matrix calculation section 335 generates a rotational component conversion matrix information on the transformation matrix F _R of the rotation component is output together with the coordinate conversion information in the translational component calculation unit 332.

As illustrated in FIG. 31, the translation component calculation unit 331 includes a matrix calculation unit 338 and an F matrix calculation unit 339. The matrix calculation unit 338 performs the same process as the matrix calculation unit 320 described above. However, the matrix calculating unit 338 calculates the operation matrix _{M 5} represented by the following formula (53).

Here, Σ means summing up for all the representative point pairs, and the matrix C ₃ is a combination of a part of the orthogonal vector group F ′ ₀ described above, and is expressed by the following equation (54). .

The matrix calculation unit 336 generates calculation matrix information regarding the calculated calculation matrix M ₅ and outputs the calculation matrix information to the F matrix calculation unit 339. As shown in equation (53), the operation matrix M _5, contains transformation matrix F _R of the rotary component. That is, operation matrix M ₅ is adapted to that in consideration of the matrix F _R of the rotary component. The conversion matrix F _R, since it is constrained in the transformation matrix F _P of the projection components, operation matrix M ₅ is essentially has to those considering the transformation matrix F _P of the projection component.

The F matrix calculation unit 339 performs the same processing as the F matrix calculation unit 321 described above based on the calculation matrix information. Thereby, a translation matrix F _Sh of the translation component shown by the following formula (55) is calculated. Here, operation matrix M ₅ is a transformation matrix F _P of the projection component and a rotational _component, F since become those considering _R, F matrix calculation unit 339, the projection component and the transformation matrix F _P of the rotating component, it is possible to calculate a transformation matrix F _Sh constrained to F _R.

Here, v1 to v4 indicate arbitrary numerical values. The F matrix calculation unit 339 calculates the transformation matrix F by multiplying the transformation matrices F _P , F _R , and F _Sh of the projection component, the rotation component, and the translation component. Then, transformation matrix information related to the calculated transformation matrix F is generated and output to the F matrix decomposition unit 322 together with the coordinate transformation information.

  F matrix decomposition section 333 shown in FIG. 28 performs the same processing as F matrix decomposition section 322 described above.

  Since the linear geometric parameter fitting unit 326 according to this modification calculates the transformation matrix F for each component, the transformation matrix F can be calculated robustly and with high accuracy. In addition, since the linear geometric parameter fitting unit 326 considers the components already calculated when calculating each component (used as a constraint condition), the transformation matrix F can be calculated with high accuracy also in this respect.

[Second Modification of Linear Geometric Parameter Fitting Unit]
Next, a second modification of the linear geometric parameter fitting unit will be described. As illustrated in FIG. 32, the linear geometric parameter fitting unit 340 according to the second modification includes an initialization unit 341, a coordinate conversion unit 342, and a linear geometric parameter update unit 343. The linear geometric parameter fitting 340 is generally different from the linear geometric parameter fitting unit 303 in that the process of calculating the transformation matrix F is simplified.

  The initialization unit 341 and the coordinate conversion unit 342 perform the same processing as the initialization unit 312 and the coordinate conversion unit 319 described above.

  The linear geometric parameter update unit 343 stores a translation component update table, a rotation component update table, and a projection component update table.

In the translation component update table, values from -5 to +5 are stored in increments of 0.5. As shown in FIG. 33, the rotation component update table and the projection component update table have a 5 × 5 matrix shape. Each block of the rotation component update table and the projection component update table stores a combination of parameters (Ta, Tb) necessary for converting the color calibrated images V _L-1 and V _R-1. Yes. Here, Ta is 0, S, or J, and Tb is 0, s, or j. “S” and “S” indicate values near 0 (steps), and “j” and “J” indicate values away from 0 (jump). That is, the difference between “s” or “S” and “j” or “J” is larger than the difference between 0 and “s” or “S”.

Specifically, in the rotation component update table, Ta indicates the rotation angle of the color calibrated image V _L-1 , and Tb indicates the rotation angle of the color calibrated image VR _-1 . That is, “S” indicates, for example, that the color calibrated image V _L-1 is rotated by 0.01 degrees in the clockwise direction, and “J” is, for example, that the color calibrated image V _L-1 is rotated to the right. Indicates a 0.1 degree rotation in the direction. "-" Indicates left rotation.

Similarly, “s” indicates, for example, that the color calibrated image V _R-1 is rotated by 0.01 degrees in the right rotation direction, and “j” indicates, for example, the color calibrated image V _R-1 image. It shows that it rotates 0.1 degree in the clockwise direction. "-" Indicates left rotation.

On the other hand, in the projection component update table, Ta indicates the size for projective conversion of the color calibrated image _VL-1 , and Tb indicates the size for projective conversion of the color calibrated image VR _-1 . That is, “S” indicates a combination of Pj1 and Pj2 in the above-described formula (49) or a value corresponding thereto, and “J” also indicates a combination of Pj1 and Pj2 in the above-described formula (49) or a corresponding value. Show. However, the values of Pj1 and Pj2 indicated by “J” are larger than the values of Pj1 and Pj2 indicated by “S”.

  Similarly, “s” represents a combination of Pj3 and Pj4 in the above-described formula (49) or a value corresponding thereto, and “j” is a combination of Pj3 and Pj4 in the above-described formula (49) or a value corresponding thereto. Indicates. However, the values of Pj3 and Pj4 indicated by “j” are larger than the values of Pj3 and Pj4 indicated by “s”.

  If all the values of the components of the homography matrices HL and HR, that is, the translation component, the rotation component, and the projection component are stored in the table, the size of the table explodes. Therefore, in the present embodiment, the size of each table is limited as described above. It should be noted that the above sizes are merely examples, and it is needless to say that each table may have other sizes.

  The linear geometric parameter updating unit 343 uses the following formulas (56) and (57) for all combinations of values in the translation component update table, rotation component update table, and projection component update table. Homography update matrices DL and DR represented by

Here, v _L and v _R indicate values stored in the translation component update table, respectively. r _L represents Ta of the rotation component update table, and v _R represents Tb of the rotation component update table. p _L indicates Ta in the projection component update table, and p _R indicates Tb in the projection component update table.

  Then, the linear geometric parameter update unit 343 updates the homography matrices HL and HR based on the following equations (58) and (59).

  Then, the linear geometric parameter updating unit 343 calculates error values for all the calculated homography matrices HL and HR using the above equation (32), and the homography matrices HL and HR having the smallest error values. Select. The linear geometric parameter update unit 343 generates linear correction coefficient information regarding the selected homography matrices HL and HR, and outputs the linear correction coefficient information to the nonlinear geometric parameter fitting unit 304 and the parameter selection unit 305.

  FIG. 34 shows how the homography matrices HL and HR converge. A curve L8 shows how the error value of Equation (33) fluctuates with the passage of time, and a curve L9 shows that the error value when only the translational components of the homography matrices HL and HR are updated changes with the passage of time. The curve L10 shows how the error value when the translational component and the rotational component of the homography matrices HL and HR are updated varies with time. As shown in FIG. 34, the homography matrices HL and HR converge to very small values by updating the translation and rotation components of the homography matrices HL and HR.

  FIG. 35 shows values when the homography matrices HL and HR converge. The bar graph L11 indicates a value where the error value of the equation (33) has converged, and the bar graph L12 indicates a value where the error value has converged when only the translational components of the homography matrices HL and HR are updated. A bar graph L13 indicates a value obtained by converging error values when the translation component and the rotation component of the homography matrices HL and HR are updated, and a bar graph L14 indicates a translation component, a rotation component, and a projection component of the homography matrices HL and HR. Indicates the value that the error value has converged when. The threshold value Th1 indicates the upper limit of the error value when the parallax detection described later is performed satisfactorily. That is, the error value needs to be equal to or less than the threshold value Th1 in order to perform the parallax detection described later well. As shown in FIG. 35, the parallax detection is favorably performed by updating at least the translation component and the rotation component of the homography matrices HL and HR.

  The linear geometric parameter fitting unit 340 according to the second modification calculates the transformation matrix F using the update table. For this reason, the linear geometric parameter fitting unit 340 can reduce the processing load when calculating the transformation matrix F. Further, since the processing load is reduced, the variation in parameters is suppressed, so that the linear geometric parameter fitting unit 340 can calculate a robust transformation matrix F.

  In addition, since the linear geometric parameter fitting unit 340 limits the size of the update table, the size of the update table can be prevented from exploding.

[Modification of parameter selection unit]
Next, a modified example (conversion matrix adjustment unit) of the parameter selection unit 305 will be described. The parameter selection unit 305 according to this modified example performs the following image matching process after selecting the homography matrices HL and HR by the above-described process.

That is, as shown in FIG. 36, the parameter selection unit 305 calculates the transformation matrix F based on the homography matrices HL and HR, transforms the left representative point with the transformation matrix F, and converts the epipolar line EP _R-1 2. And The parameter selection unit 305, a perpendicular line is drawn from right representative point _{P R-1} 2 the epipolar line _{EP R-1} 2, obtaining the intersection point _{P R-1} 3. Then, the parameter selection unit 305 includes a first peripheral patch (for example, a 10 × 10 pixel patch) Pat1 centered on the right representative point P _R- 12 and a second peripheral centered on the intersection P _R-13. A patch (for example, a patch of 10 × 10 pixels) Pat2 is extracted. Then, the parameter selection unit 305 adjusts the homography matrices HL and HR so that the correlation between these patches is maximized. Specifically, the parameter selection unit 305 adjusts the homography matrices HL and HR so that the value of the following equation (60) is minimized.

Here, a _j and a _i indicate the luminance of each pixel in the first patch Pat1, and b _i and b _k indicate the luminance of each pixel in the second patch Pat2.

  Then, the parameter selection unit 305 repeats the process described above using the adjusted homography matrices HL and HR and other representative pairs. The parameter selection unit 305 performs the above-described processing for all representative pairs, generates parameter information regarding the homography matrices HL and HR, and outputs the parameter information to the linear geometric parameter fitting unit 303 and the geometric calibration execution unit 306. To do.

  Therefore, the parameter selection unit 305 can calculate the homography matrices HL and HR, and hence the transformation matrix F, with higher accuracy.

  In addition, you may combine the modification mentioned above arbitrarily. For example, all the modified examples may be combined. In this case, the linear geometric parameter fitting unit 303 generates homography matrices HL and HR by a plurality of methods. Then, the parameter selection unit 305 selects an optimal one of the homography matrices HL and HR generated by the homography matrices HL and HR and the nonlinear geometric parameter fitting unit 304, and further, the third modification described above. The processing of the example improves the accuracy of the homography matrices HL and HR. The homography matrices HL and HR are used as initial parameters in the calculation in the next frame.

<2-3. Configuration of parallax detection unit>
Next, the configuration of the parallax detection unit 4 illustrated in FIG. 5 will be described. As shown in FIG. 37, the parallax detection unit 4 includes a global matching unit 401, a local matching unit 402, and a disparity map integration unit 403. That is, the parallax detection unit 4 performs global matching and local matching in parallel. Here, the local matching results and the model of the calibration accuracy of the quality is above (for example, a model formula showing the relationship between the luminance of the input image V _R of the input image V _L, the transform matrix F) to It has the advantage of not relying on it, but it also has the disadvantages of being vulnerable to occlusion and poor stability (accuracy tends to vary). On the other hand, global matching has the advantage of being strong and stable in occlusion, but has the disadvantage that accuracy is likely to depend on the calibration result and model. Therefore, the parallax detection unit 4 performs both in parallel and compares the resulting parallax maps to select a good parallax map.

[Configuration of global matching section]
As shown in FIG. 38, the global matching unit 401 (first horizontal parallax detection unit) includes a DSAD (Dynamic Sum Of Absolute Difference) calculation unit 404, a minimum value selection unit 405, a motion description unit 406, an anchor vector, A construction unit 407, a cost calculation unit 408, a route construction unit 409, and a backtrack unit 410 are provided.

[Configuration of DSAD calculation unit]
The DSAD calculation unit 404 performs the following processing for each left coordinate point based on the calibrated images V _L-2 and VR _-2 and the reliability map information given from the color calibration unit 2. That is, the DSAD calculation unit 404 extracts all the right coordinate points having the same height (same y coordinate) as the left coordinate points as reference coordinate points, and the left coordinate points as reference coordinate points. Then, the DSAD calculation unit 404 subtracts the x coordinate of the standard coordinate point from the x coordinate of each reference coordinate point, and sets the value obtained thereby as the horizontal parallax d1 of the standard coordinate point. The DSAD calculation unit 404 then calculates DSAD (j) (j is an integer of −2 to 2) represented by the following formula (61) for all horizontal parallaxes d1 (that is, all reference coordinate points) of the standard coordinate points. ) Is calculated.

  Here, i is an integer of −2 to 2, L (i, j) is the luminance of the left coordinate point different from the reference coordinate point by (i + j) and R (i, j) is a reference. This is the luminance at the right coordinate point where the coordinate point differs from the y coordinate by (i + j). Further, α1 is obtained by normalizing the reliability of the reference coordinate point indicated by the color calibration reliability map to a value of 0 to 1, and becomes closer to 1 as the reliability increases. Thus, by using the value of the reliability map for color calibration for the calculation of DSAD, it is possible to detect parallax that is resistant to color misregistration. Normalization is obtained by substituting the reliability into the above-described equation (16).

  Therefore, the DSAD calculation unit 404 calculates several DSADs by shaking the y-coordinates of the standard coordinate point and the reference coordinate point. Thereby, the parallax detection part 3 can produce | generate the parallax map which considered the vertical shift | offset | difference of the left coordinate point and the right coordinate point. The DSAD calculation unit 404 sets the y-coordinate amount of the reference coordinate point to two pixels above and below the y-coordinate of the standard coordinate point, but this range can be arbitrarily changed according to the balance between robustness and accuracy Is done.

  The DSAD calculation unit 404 generates DSAD information regarding the calculated DSAD (−2) to DSAD (+2) and outputs the DSAD information to the minimum value selection unit 405.

[Configuration of minimum value selector]
The minimum value selection unit 405 selects the minimum DSAD for all the left coordinate points and the horizontal parallax d1 based on the DSAD information. Then, the minimum value selection unit 405 stores the selected DSAD in each node P (x, d) of the DP map for parallax detection shown in FIG. Therefore, the minimum DSAD is the score of the node P (x, d). In the DP map for parallax detection, the horizontal axis is the x coordinate of the left coordinate point, and the vertical axis is the horizontal parallax d1, and has a plurality of nodes P (x, d). The DP map for parallax detection is used when calculating the horizontal parallax d1 of the left coordinate point. Further, the DP map for parallax detection is generated for each y coordinate of the left coordinate point. Therefore, any node P (x, d) in any DP map for parallax detection corresponds to any left coordinate point.

[Composition of motion description part]
As shown in FIG. 39, the motion description unit 406 includes a horizontal difference calculation unit 411, a vertical difference calculation unit 412, a time difference calculation unit 413, a filter unit 414, and a motion score calculation unit 415.

Horizontal difference calculation unit 411 obtains the calibrated image V _L-2, the pixels of the calibrated image V _L-2, i.e., for each left coordinate point, the following process is performed. That is, the horizontal difference calculation unit 411 uses the left coordinate point as the reference coordinate point, and subtracts the luminance of the reference coordinate point whose x coordinate is one greater than the reference coordinate point from the luminance of the standard coordinate point. The horizontal difference calculation unit 411 sets the obtained value as the horizontal difference dx, generates horizontal difference information regarding the horizontal difference dx, and outputs the horizontal difference information to the filter unit 414.

Vertical difference calculating section 412 obtains the calibrated image V _L-2, the pixels of the calibrated image V _L-2, i.e., for each left coordinate point, the following process is performed. That is, the vertical difference calculation unit 412 uses the left coordinate point as the reference coordinate point, and subtracts the luminance of the reference coordinate point whose y coordinate is one greater than the reference coordinate point from the luminance of the standard coordinate point. The vertical difference calculation unit 412 sets the obtained value as the vertical difference dy, generates vertical difference information regarding the calculated vertical difference dy, and outputs the vertical difference information to the filter unit 414.

The time difference calculation unit 413 acquires the calibrated image V _L-2 and the calibrated image V ′ _{L-2 of} the previous frame, and each pixel of the calibrated image V _L-2 , that is, for each left coordinate point. In addition, the following processing is performed. That is, the time difference calculation unit 413 uses the left coordinate point as the reference coordinate point, and among the pixels of the calibrated image V ′ _L-2 in the previous frame, the pixel having the same coordinate as the reference coordinate point is set as the reference coordinate point. To do. Then, the time difference calculation unit 413 subtracts the luminance of the reference coordinate point from the luminance of the standard coordinate point. The time difference calculation unit 413 generates the time difference information related to the calculated time difference dt by using the obtained value as the time difference dt, and outputs the time difference information to the filter unit 414.

  The filter unit 414 applies a low-pass filter to the horizontal difference information, the vertical difference information, and the time difference information. Thereafter, the filter unit 414 outputs these pieces of information to the motion score calculation unit 415.

  The motion score calculation unit 415 performs static motion determination for each left coordinate point based on the horizontal difference information, the vertical difference information, and the time difference information, and determines a motion score according to the result.

  Specifically, the motion score calculation unit 415 calculates a time differential value (dx / dt + dy / dt) (luminance change amount per unit time) of the luminance change amount for each left coordinate point, and calculates the calculated differential value. Is normalized to a value of 0 to 1 by substituting into the sigmoid function. The motion score calculation unit 415 uses the normalized value as the motion score, and outputs motion score information related to the motion score to the anchor vector construction unit 407. The motion score becomes smaller as the luminance variation of the left coordinate point becomes more intense (that is, the motion of the image becomes more intense).

  The motion score calculation unit 415 may calculate the motion score as follows.

That is, the motion score calculation unit 415 generates the mosaic image V ″ _L-2 by reducing the resolution of the calibrated image V _L-2 . The motion score calculation unit 415 uses the left coordinate point of the calibrated image V _L-2 as a reference coordinate point, and selects a coordinate point having the same coordinates as the reference coordinate point in the mosaic image V ″ _L-2 as a reference coordinate point. And The motion score calculation unit 415 subtracts the luminance of the reference coordinate point from the luminance of the standard coordinate point. The motion score calculation unit 415 normalizes the calculated difference value to a value of 0 to 1 by substituting the calculated difference value into the sigmoid function. The motion score calculation unit 415 uses the normalized value as the motion score. In this case, the motion score increases in a plain portion of the calibrated image V _L-2 . With this method, the processing load is reduced and a robust motion score can be obtained.

[Configuration of Anchor Vector Building Unit]
The anchor vector construction unit 407 illustrated in FIG. 38 includes a local histogram creation unit 416 and an anchor vector calculation unit 417 as illustrated in FIG.

The local histogram creation unit 416 acquires the parallax map DM _n−1 created in the previous frame. The local histogram creation unit 416 performs the following process for each coordinate point of the parallax map (these coordinate points correspond to the left coordinate point). Specifically, the local histogram creation unit 416 uses any coordinate point of the parallax map as a reference coordinate point and starts from a peripheral area of the reference coordinate point (for example, a 5 × 5 pixel area centered on the reference coordinate point). Horizontal parallax information is extracted. Then, the local histogram creation unit 416 generates a local histogram DH shown in FIG. 41 based on the horizontal parallax information. The local histogram DH shows a correspondence relationship between the value of the horizontal parallax d1 and the number (frequency) of pixels having the parallax value.

  The local histogram creation unit 416 generates local histogram information regarding the local histogram DH and outputs the local histogram information to the anchor vector calculation unit 417.

  The anchor vector calculation unit 417 performs the following processing for each left coordinate point based on the local histogram information. That is, as shown in FIG. 41, the anchor vector calculation unit 417 searches for a horizontal parallax d1, that is, a high-frequency parallax d1 'whose frequency is equal to or higher than a predetermined threshold Th2. The anchor vector calculation unit 417 generates anchor vector information indicating a bonus value obtained by multiplying the motion score by a predetermined value and the high-frequency parallax d1 ', and outputs the anchor vector information to the cost calculation unit 408. The anchor vector information is expressed by, for example, the following formula (67).

Here, α2 indicates a bonus value, and the matrix M _d indicates the position of the high-frequency parallax d1 ′. That is, each column of the matrix M _d indicates a different horizontal parallax d1, and a column whose component is 1 indicates that the horizontal parallax d1 corresponding to the column is a high-frequency parallax d1 ′. When the high frequency parallax d1 ′ does not exist, all the components of the matrix M _d become zero.

[Configuration of cost calculation unit]
The cost calculation unit 408 illustrated in FIG. 38 updates the value of each node P (x, d) in the DP map for parallax detection based on the anchor vector information. That is, the cost calculation unit 408 identifies a node (x, d (= d1 ′)) corresponding to the high frequency parallax d1 ′ for each left coordinate point, and subtracts the bonus value α2 from the score of this node. Thereby, a node having the same parallax as the parallax having a high frequency in the previous frame is likely to pass through the shortest path. In other words, parallax having a high frequency in the previous frame is easily selected in the current frame.

[Configuration of route building unit]
The path construction unit 409 illustrated in FIG. 38 includes a left-eye image horizontal difference calculation unit 418, a right-eye image horizontal difference calculation unit 419, a weight calculation unit 420, and a path calculation unit 421 as illustrated in FIG. .

  The left-eye image horizontal difference calculation unit 418 performs the same processing as the horizontal difference calculation unit 411 described above. Then, the left-eye image horizontal difference calculation unit 418 outputs the horizontal difference information generated by this processing to the weight calculation unit 420.

The right-eye image horizontal difference calculation unit 419 acquires the calibrated image VR _-2 . The right-eye image horizontal difference calculation unit 419 performs the same process as the above _- described horizontal difference calculation unit 411 on the calibrated image VR _-2 . Then, the right-eye image horizontal difference calculation unit 419 outputs the horizontal difference information generated by this processing to the weight calculation unit 420.

The weight calculation unit 420 calculates the weights wt _{L of} the left coordinate points and wt _{R of the} right coordinate points for all coordinate points based on the horizontal difference information. Specifically, the weight calculation unit 420 normalizes the horizontal difference dw _L to a value of 0 to 1 by substituting the horizontal difference dw _L of the left coordinate point into a sigmoid function, and sets this as the weight wt _L. Similarly, the weight calculation unit 420 normalizes the horizontal difference dw _R to a value of 0 to ₁ by substituting the horizontal difference dw _R of the right coordinate point into the sigmoid function, and sets this as the weight wt _R. Then, the weight calculation unit 420 generates weight information regarding the calculated weights wt _L and wt _R , and outputs the weight information to the route calculation unit 421. The weights wt _L and wt _R become smaller at the edge portion (outline) of the image and become larger at the flat portion.

The route calculation unit 421 calculates an accumulated cost up to each node P (x, d) of the DP map for parallax detection based on the weight information given from the weight calculation unit 420. Specifically, the route calculation unit 421 uses the node (0, 0) as the start point, the node (x _maz , 0) as the end point, and the accumulated cost from the start point to the node P (x, d) is as follows: Defined in Here, x _max is the maximum value of the x coordinate of the left coordinate point.

Here, DFI (x, d) ₀ is the accumulated cost when reaching the node P (x, d) through the path PA _d 0, and DFI (x, d) ₁ passes through the path PA _d 1. DFI (x, d) ₂ is the accumulated cost when reaching the node P (x, d) through the path PA _d 2. DFI (x, d-1) is the accumulated cost from the starting point to the node P (x, d-1). DFI (x-1, d) is the accumulated cost from the starting point to the node P (x-1, d). DFI (x-1, d + 1) is an accumulated cost from the start point to the node P (x-1, d + 1). Further, occCost ₀ and occCost ₁ are predetermined values indicating cost values, for example, 4.0. wt _L is the weight of the left coordinate point corresponding to the node P (x, d), and wt _R is the weight of the right coordinate point having the same coordinates as the left coordinate point.

Then, the route calculation unit 421 selects the smallest one of the calculated accumulated costs DFI (x, d) _{0 to} DFI (x, d) ₂ , and selects the selected one as the node P (x, d). Is the accumulated cost DFI (x, d). The route calculation unit 421 calculates the accumulated cost DFI (x, d) for all the nodes P (x, d) and stores it in the DP map for parallax detection.

  The backtrack unit 410 calculates the shortest path, that is, the path with the minimum accumulated cost from the start point to the end point by tracing back the path with the minimum accumulated cost from the end point to the start point. The node on this shortest path is the horizontal parallax d1 of the corresponding left coordinate point.

Backtracking unit 410, the shortest path calculated, based on the vertical parallax information supplied from the local matching unit to be described later, to generate a disparity map DM _1. Parallax map DM ₁ shows a horizontal parallax d1 and the vertical disparity d2 for each left coordinate point. Backtracking portion 410 generates a disparity map information about the parallax map DM ₁ generated, and outputs the disparity map integration unit 403 shown in FIG. 37. Here, the parallax map DM ₁ is optimally selected by the local matching unit 402 out of the difference between the y coordinate of the left coordinate point and the y coordinate of the right coordinate point having the same x coordinate as the left coordinate point, that is, the vertical parallax d2. What is determined to be is stored.

[Configuration of local matching unit]
Local matching section 402 (second horizontal parallax detection unit) by performing a so-called block matching, calculates a disparity map DM ₂ for each pixel block. Parallax map DM ₂ also shows a horizontal parallax d1 and the vertical disparity d2 for each left coordinate point.

Specifically, the local matching unit 402 divides the calibrated images V _L-2 and V _R-2 into a plurality of pixel blocks, respectively. For example, the local matching unit 402 divides the calibrated images V _L−2 and V _R−2 into 64 pixel blocks, like the geometric calibration unit 3 described above. Then, the local matching unit 402 performs the same processing as that of the global matching unit 401 described above for each pixel block (hereinafter, also referred to as “left pixel block”) of the calibrated image V _L-2 . Here, the local matching unit 402 also calculates the vertical parallax d2 for each left coordinate point. The calculation method of the vertical parallax d2 is the same as the processing of the global matching unit 401 described above. For example, the local matching unit 402 calculates the DSAD in the y-axis direction by using the right coordinate point having the same x coordinate as the reference coordinate point as the reference coordinate point in the above-described process of calculating the DSAD. Thereafter, the local matching unit 402 generates a parallax detection DP map for calculating the vertical parallax d2 based on the calculated DSAD and the like, calculates the shortest path of the parallax detection DP map, and the shortest path Based on the above, the vertical parallax d2 is calculated for each left coordinate point.

Local matching unit 402, while outputting the disparity map information about the parallax map DM ₂ generated disparity map integration unit 403 generates the vertical parallax information about vertical parallax d2, and outputs to the global matching unit 401.

[Configuration of Disparity Map Integration Unit]
First, the disparity map integration unit 403 evaluates the disparity map DM ₁ generated by the global matching unit 401 and the disparity map DM ₂ generated by the local matching unit 402 based on the disparity map information.

Specifically, the disparity map integration unit 403 generates a first difference map indicating a difference between the disparity map DM ₁ of the current frame and the disparity map DM ₁ of the previous frame. First difference map, a value obtained by subtracting the horizontal disparity d1 disparity map DM ₁ of the previous frame from the horizontal parallax d1 disparity map DM ₁ of the current frame, showing for each left coordinate point. Next, the disparity map integration unit 403 generates a first binarized difference map by binarizing the first difference map. Then, the disparity map integration unit 402 generates a first difference score map by multiplying each value of the first binarized difference map by a predetermined weight (for example, 8).

Further, the disparity map integration unit 403, a disparity map DM ₁ of the current frame, and generates an edge image of the calibrated image V _L-2 of the current frame to generate a correlation map showing these correlations. Parallax map DM ₁ edge image of the current frame indicates a disparity map DM ₁ of the edge portion of the current frame (contour portion of each image drawn on parallax map DM ₁ of the current frame). Similarly, calibrated image V _L-2 of the edge image shows a calibrated image V _L-2 of the edge portion (contour portion of each image drawn on the calibrated image V _L-2). As a method for calculating the correlation between the edge images, for example, a method for calculating the correlation such as NCC is arbitrarily used. Then, the disparity map integration unit 403 generates a binarized correlation map by binarizing the correlation map. Then, the disparity map integration unit 402 generates a correlation score map by multiplying each value of the binarized correlation map by a predetermined weight (for example, 26).

The disparity map integration unit 403, a correlation score map and the first difference score map integrates, by multiplying the IIR filter to generate a global matching reliability map EM _G. The value of each left coordinate point of global matching reliability map EM _G, of the value of the first difference score map and the correlation score map, showing the larger value.

On the other hand, disparity map integration unit 403 generates a disparity map DM ₂ of the current frame, the second difference map indicating the difference between the disparity map DM ₂ of the previous frame. Difference map, a value obtained by subtracting the horizontal disparity d1 disparity map DM ₁ of the previous frame from the horizontal parallax d1 disparity map DM ₁ of the current frame, showing for each left coordinate point. Next, the disparity map integration unit 403 generates a second binarized difference map by binarizing the second difference map. Then, the disparity map integration unit 402 generates a second difference score map by multiplying each value of the second binarized difference map by a predetermined weight (for example, 16).

Further, the disparity map integration unit 403 generates an edge image of the calibrated image _VL-2 of the current frame. This edge image shows a calibrated image V _L-2 of the edge portion (contour portion of each image drawn on the calibrated image V _L-2). The disparity map integration unit 403 generates a binarized edge map by binarizing the edge image. Then, the disparity map integration unit 402 generates an edge score map by multiplying each value of the binarized edge map by a predetermined weight (for example, 8).

The disparity map integration unit 403, the second integrating and differential score map and the edge score map, by multiplying the IIR filter, to generate the local matching reliability map EM _L. The value of each left coordinate point of the local matching reliability map EM _L, of the value of the second difference score map and the edge score map, showing the larger value.

Thus, the disparity map integration unit 403, and evaluated by different evaluation methods parallax map DM ₁ generated by the global matching unit 401, by integrating the results, the global matching reliability map EM _G Generate. Similarly, the disparity map integration unit 403, and evaluated by different evaluation methods a disparity map DM ₂ generated by the local matching unit 402, by integrating the results, generates a local matching reliability map EM _L To do. Here, it is different from each other and evaluation methods for evaluation of the disparity map DM ₁ approaches the parallax map DM _2. Different weighting is performed by the evaluation method.

The disparity map integration unit 403, a global matching reliability map EM _G, by comparing the local matching reliability map EM _L, each left coordinate point, the disparity map DM ₁ and parallax map DM ₂ It is determined which is more reliable. Based on the determination result, the disparity map integration unit 403 generates a parallax detection reliability map EM indicating a parallax map with high reliability for each left coordinate point.

FIG. 44 shows an example of the parallax detection reliability map EM. That is, FIG. 44A shows the reliability map EM1 for parallax detection generated when the geometric calibration is not performed, and FIG. 44B is generated when the geometric calibration is performed. The parallax detection reliability map EM2 is shown. Region EM11 is towards the disparity map DM ₁ indicates a region reliable than parallax map DM _2, region EM12 is towards the disparity map DM ₂ indicates a region higher reliability than the disparity map DM _1.

  Since local matching does not depend on the result of calibration as described above, the region EM12 in the parallax detection reliability map EM1 is wider than the region EM12 in the parallax detection reliability map EM2.

  On the other hand, since the global matching can perform good parallax detection when accurate geometric calibration is performed as in the present embodiment, the region EM11 in the parallax detection reliability map EM2 can be used. Is wider than the region EM11 in the parallax detection reliability map EM1.

The disparity map integration unit 403 generates a disparity map DM based on the generated disparity detection reliability map. Here, the horizontal parallax d1 of each left coordinate point of the disparity map DM indicates the value of the higher reliability of the parallax map DM ₁ and disparity maps DM _2.

FIG. 45 is a diagram for explaining the effect of the parallax detection 3. FIG. 45A shows a multi-viewpoint image V _V 1 generated based on the parallax map DM ₁ , and FIG. 45B shows a multi-viewpoint image V _V 2 generated based on the parallax map DM _2. FIG. 45 (c) shows a multi-viewpoint image V _V 3 generated based on the parallax map DM. In the multi-view images V _V 1 to _{V V} 3, subject images V _V 10 to _{V V} 30 are drawn. The tip portions V _V 11 to _{V V} 31 of the subject images V _V 10 to _{V V} 30 are regions to which the above-described model is not applied.

In global matching, the accuracy of accuracy depends on the model, so the end V _V 11 is broken. However, since the accuracy of local matching does not depend on the model, the end V _V 21 is not broken. Since the parallax map DM indicates a value with higher reliability of the parallax map DM ₁ and the parallax map DM ₂ , the value of the parallax map DM ₂ is indicated in the region of the end V _V 31. Therefore, the end portion V _V 31 is generated based on the parallax map DM ₂ . For this reason, the end portion V _V 31 is not broken.

[Effects of parallax detector]
Next, effects of the parallax detection unit 4 will be described.

  Since the parallax detection unit 4 considers parallax in the y-axis direction when calculating the DSAD, the parallax detection unit 4 can perform parallax detection that is resistant to geometrical deviation (vertical deviation). That is, the parallax detection unit 4 can perform horizontal parallax detection that is more robust and more systematic than conventional ones. Moreover, since the parallax detection unit 4 limits the range in which the parallax in the y-axis direction is considered (here, two pixels above and below the reference coordinate point), it is possible to prevent the explosion of the optimization problem from occurring. .

  In addition, since the parallax detection unit 4 considers the reliability calculated in the color calibration when calculating the DSAD, the parallax detection unit 4 can perform parallax detection in consideration of the balance between robustness and accuracy.

  Further, since the parallax detection unit 4 uses the DSAD as the score of the parallax detection DP map, it can calculate the score of the parallax detection DP map with higher accuracy than when only the SAD is used as a score, Parallax detection can be performed with high accuracy.

Moreover, since the parallax detection unit 4 gives a bonus to each score of the parallax detection DP map based on the result of the static motion determination, the score of the parallax detection DP map can be calculated with high accuracy. Also, the disparity detection unit 4 'generates _L-2, the mosaic image V' mosaic image V 'by lowering the resolution of the calibrated image V _L-2 to calculate the motion score based on' _L-2 May be. In this case, the processing load is reduced and a robust motion score can be obtained.

Moreover, since the parallax detection unit 4 considers the weights wt _L and wt _R according to the horizontal difference when calculating the accumulated cost of each node P (x, d), the accumulated cost can be calculated with high accuracy. it can. Since the weights wt _L and wt _R are small at the edge portion and large at the flat portion, smoothing is appropriately performed according to the image.

Also, the disparity detection unit 4, the disparity map DM ₁ calculated by the global matching, since the inclusion of vertical parallax d2 calculated by the local matching, it is possible to improve the accuracy of the parallax map DM _1. For example, the disparity map DM _1, the input image _V L, even if no 0.00 alignment _{V R,} becomes robust.

Further, the parallax detection unit 4 generates a correlation map indicating the correlation between the edge images of the parallax map DM ₁ and the calibrated image V _L-2, and calculates the reliability of the parallax map DM ₁ based on the correlation map. since, it is possible to calculate the reliability of the so-called streaking area disparity map DM _1. For this reason, the parallax detection unit 4 can accurately perform parallax detection in the streaking area.

Also, the disparity detection unit 4, when evaluating the parallax map DM ₁ and parallax map DM _2, so to evaluate the parallax map DM ₁ and parallax map DM ₂ at different evaluation method, considering these characteristics Evaluation can be performed.

Also, the disparity detection unit 4, by multiplying the IIR filter to map obtained by the respective evaluation methods, and the global matching reliability map EM _G, because it produces a local matching reliability map EM _L, time A stable reliability map can be generated.

Also, the disparity detection unit 4, of the parallax map DM ₁ and parallax map DM _2, because it generates a disparity map DM with higher reliability, hard regions accurate parallax detected in the global matching and, Accurate parallax can be detected in a region where accurate parallax is difficult to detect by local matching.

  In addition, since the parallax detection unit 4 considers the generated parallax map DM in the next frame, the parallax detection can be performed with higher accuracy than when a plurality of matching methods are simply performed in parallel.

<3. Processing by Image Processing Device>
Next, processing by the image processing apparatus 1 will be described.

[Processing by color calibration unit]
First, processing performed by the color calibration unit 2 will be described with reference to the flowchart shown in FIG.

In step S10, the block dividing unit 207 divides the input image V _L into eight pixel blocks B _{L−1 as shown} in FIG. Similarly, the block dividing unit 207, as shown in FIG. 8 (b), divides the input image _{V R} into eight pixel blocks _{B R-1.}

Next, the block dividing unit 207 reduces the gradation in each pixel block B _L-1 and _BR-1 to 64 gradations. The block dividing unit 207 outputs block information regarding the pixel blocks B _L-1 and B _R-1 to the histogram generating unit 208.

Next, the histogram generation unit 208 generates a color histogram CH _L as shown in FIG. 9A for each pixel block B _L−1 based on the block information. Similarly, the histogram generation unit 208 generates a color histogram CH _R as shown in FIG. 9B for each pixel block _BR-1 . Next, the histogram generation unit 208 associates the color histogram CH _L with the color histogram CH _R. Next, the histogram generation unit 208 generates color histogram information regarding the color histograms CH _L and CH _R and outputs the color histogram information to the DP matching unit 209.

Next, the DP matching unit 209 performs the following processing for each set of the pixel blocks B _L-1 and B _R-1 based on the color histogram information. That is, the DP matching unit 209 first generates a color correspondence detection DP map as shown in FIG. Then, the DP matching unit 209 uses the node P (0, 0) as the start point, the node P (63, 63) as the end point, and the accumulated cost from the start point to the node P (n _L , n _R ) It is defined as (1) to (3).

Next, the DP matching unit 209 calculates the accumulated cost from the start point to each node including the end point, and by tracing back the route having the minimum accumulated cost from the end point to the start point, the shortest route, that is, the start point to the end point. The route with the minimum accumulated cost is calculated. The DP matching unit 209 generates the shortest path information regarding the shortest path calculated for each set of the pixel blocks B _L-1 and B _R-1 and outputs the shortest path information to the color pair calculation unit 210.

Then, color-pair calculation unit 210, based on the shortest path information, calculates the color pairs for each set of pixel blocks _{B L-1, B R-} 1. The color pair calculation unit 210 calculates about 250 color pairs per pixel block _BL-1, and calculates about 2000 color pairs in total. The color pair calculation unit 210 generates color pair information regarding the calculated color pair, and outputs the color pair information to the linear color parameter fitting unit 202 and the non-linear color parameter fitting unit 204.

In step S20, the linear color parameter fitting portion 202, on the basis of the color pair information, it calculates the model formula showing the relationship between the luminance of the input image V _R of the input image V _L. This model formula is expressed by the above-described formula (7).

Specifically, first, the addition matrix calculation unit 211 calculates the matrix represented by the above-described formula (8) for all color pairs. The addition matrix calculation unit 211 calculates the addition matrix M ₁ represented by the above-described equation (9) by adding all the calculated matrices.

Then, the matrix calculator 211 summing generates a summing matrix information on the matrix M ₁ summation is calculated, and outputs the coefficient calculation unit 212 with the color pair information.

  Next, the coefficient calculation unit 212 calculates the matrix represented by the above-described formula (10) for all color pairs based on the color pair information. The coefficient calculation unit 212 calculates the coefficient calculation matrix A represented by the above-described equation (11) by adding all the calculated matrices.

Next, the coefficient calculation unit 212 calculates the first initial value coef (w) ₁ of the color correction coefficients wi and wa based on the above-described equation (12).

Next, the coefficient calculation unit 212 generates first initial value information regarding the calculated first initial value coef (w) ₁ and outputs the first initial value information to the filter unit 213.

Next, the filter unit 213, based on the first initial value information given from the coefficient calculation unit 212, the linear correction coefficient information of the previous frame given from the coefficient holding unit 214, and the above-described equation (13), A second initial value coef (w) ₂ of the color correction coefficients wi and wa is calculated.

Next, the filter unit 213 applies the IIR filter to the second initial value coef (w) ₂ to calculate the color correction coefficients wi and wa as linear correction coefficients coef (linear). The filter unit 213 outputs the linear correction coefficient information regarding the calculated linear correction coefficient coef (linear) to the coefficient holding unit 214, the nonlinear color parameter fitting unit 204, the evaluation unit 205, and the color calibration execution unit 206.

On the other hand, feature matching unit 203, by matching the brightness of the input image V _L, V _R of the current frame, and extracts the color pair. Next, the feature amount matching unit 203 generates color pair information regarding the extracted color pair, and outputs the color pair information to the nonlinear color parameter fitting unit 204 and the evaluation unit 205.

  In step S30, the nonlinear color parameter fitting unit 204 calculates the color correction coefficients wi and wa as nonlinear correction coefficients coef (nonlinear) based on the color pair information and the linear correction coefficient information or the nonlinear correction coefficient information.

Specifically, first, the Jacobian matrix calculation unit 215 calculates the above-described formula based on the linear correction coefficient information or the nonlinear correction coefficient information, and the color pair information given from the histogram matching unit 201 and the feature amount matching unit 203. (14) for calculating the Jacobian matrix _{J 1} represented by. Then, the Jacobian matrix calculating unit 215 generates Jacobian matrix information related to the calculated Jacobian matrix J ₁ and outputs the generated Jacobian matrix information to the Hessian matrix calculating unit 216.

Next, the Hessian matrix calculation unit 216 calculates the Hessian matrix H ₁ represented by the above-described equation (17) based on the information given from the Jacobian matrix calculation unit 216. Next, the Hessian matrix calculation unit 216 generates Hessian matrix information regarding the calculated Hessian matrix H _1, and outputs it to the coefficient update unit 217 together with the information given from the Jacobian matrix calculation unit 215.

Next, the coefficient updating unit 217 calculates the color correction coefficients w _i and w _a as nonlinear correction coefficients coef (nonlinear) based on the information given from the Hessian matrix calculation unit 216 and the above-described equation (18). .

  Next, the coefficient updating unit 217 generates nonlinear correction coefficient information regarding the calculated nonlinear correction coefficient coef (nonlinear), and determines whether the nonlinear correction coefficient coef (nonlinear) has converged to a certain value. When it is determined that the nonlinear correction coefficient coef (nonlinear) has converged, the coefficient update unit 217 outputs nonlinear correction coefficient information to the evaluation unit 205 and the color calibration execution unit 206, and the nonlinear correction coefficient coef (nonlinear) If it is determined that it has not converged, the nonlinear correction coefficient information is output to the Jacobian matrix calculation unit 215. Thereafter, the process of step S30 is performed again.

In step S40, the color conversion unit 218 acquires an input image _V L, _{V R} of the current frame. Color conversion unit 218, the input image V _L, and V _R, and the linear correction coefficient information given from the linear color parameter fitting portion 202, on the basis to correct the brightness of the input image V _R. Color conversion unit 218, the value of the corrected luminance of the new input image V _R. Color conversion unit 218 outputs the histogram-based reliability calculation unit 219 and the input image _{V R} and the corrected input image _{V L} linear color calibrated image _{V L-1a,} as _{V R-1a.}

Next, the histogram-based reliability calculation unit 219 divides the linear color calibrated images V _L-1a and V _R-1a into eight pixel blocks B _L−1 and _BR−1 , respectively. Then, the histogram-based reliability calculation unit 219 generates a corrected color histogram indicating the correspondence between the luminance and the number (frequency) of pixels having the luminance for each of the pixel blocks B _L-1 and B _R-1. To do.

Next, the histogram-based reliability calculation unit 219 uses the normalized cross-correlation (NCC) for the similarity between the corrected color histogram of the pixel block _{BL-1 and} the corrected color histogram of the pixel block _BR-1. To calculate.

The histogram-based reliability calculation unit 219 calculates the similarity for each pixel block _BL-1 , and substitutes these similarities into the error of the above-described formula (16), for example, to reduce the similarity between 0 and 1. Normalize to a value within the range. Then, the histogram base reliability calculation unit 219 sets the normalized value as the reliability of the linear correction coefficient coef (linear). Then, the histogram-based reliability calculation unit 219 generates linear correction coefficient reliability information regarding the calculated reliability of each pixel block _BL-1 , and outputs the linear correction coefficient reliability information to the reliability map generation unit 221.

On the other hand, the feature point base reliability calculation unit 220 classifies the feature points in the input image V _L into eight pixel blocks B _L−1 based on the color pair information given from the feature amount matching unit 203. Then, the feature point base reliability calculation unit 220 calculates the reliability of the nonlinear correction coefficient coef (nonlinear) for each pixel block B _L−1 based on the nonlinear correction coefficient information given from the nonlinear color parameter fitting unit 204. To do. The feature point base reliability calculation unit 220 generates nonlinear correction coefficient reliability information related to the calculated reliability of each pixel block _BL-1 , and outputs it to the reliability map generation unit 221.

  Next, the reliability map generation unit 221 generates a color calibration reliability map based on the information given from the histogram base reliability calculation unit 219 and the feature point base reliability calculation unit 220. Then, the reliability map generation unit 221 generates reliability map information regarding the generated color calibration reliability map, and outputs the reliability map information to the color calibration execution unit 206 and the parallax detection unit 4.

In step S50, color calibration execution unit 206 acquires an input image _V L, _{V R} of the current frame. Further, the color calibration execution unit 206 determines which correction coefficient to use for each pixel block _BL-1 based on the reliability map information. Then, the color calibration execution unit 206 executes color calibration for each pixel block _BL-1 based on the linear correction coefficient information and the non-linear correction coefficient information.

[Processing by geometric calibration unit]
Next, a processing procedure by the geometric calibration unit 3 will be described with reference to a flowchart shown in FIG.

In step S60, the feature matching unit 301, by matching the brightness of the color calibrated image _{V L-1, V R-} 1, and calculates the feature vector VE _1. Next, the feature amount matching unit 301 generates feature amount vector information regarding the pixel blocks B _L-2 and _BR-2 and the feature amount vector VE _1, and outputs them to the corresponding point refinement unit 302 and the nonlinear geometric parameter fitting unit 304. To do.

In step S70, the histogram generation unit 307, based on the feature quantity vector information, and generates a vector component histogram showing the frequency of each component of the feature vector VE ₁ for each pixel block B _L-2. The histogram generation unit 307 generates vector component histogram information related to the vector component histogram and outputs the vector component histogram information to the filter unit 308.

  Next, the filter unit 308 smoothes the vector component histogram by applying a Gaussian filter to the vector component histogram information a plurality of times. The filter unit 308 generates smoothed histogram information related to the smoothed vector component histogram, and outputs the smoothed histogram information to the maximum frequency detection unit 309 and the corresponding point selection unit 311.

Next, the maximum frequency detection unit 309 specifies an x component and a y component having the maximum frequency from each vector component histogram based on the smoothed histogram information. The feature vector VE ₁ having such x component and y component is set as the representative vector VE _{2 of the} pixel block _BL-2 to which the feature vector VE ₁ belongs. Thus, the maximum frequency detecting unit 309 calculates a representative vector VE ₂ for each pixel block _{B L-2.} The maximum frequency detection unit 309 generates representative vector component information regarding the component of the representative vector VE ₂ and outputs the representative vector component information to the coordinate assigning unit 310.

Next, the coordinate assigning unit 310 determines the coordinates (start point coordinates) of each representative vector VE ₂ based on the representative vector component information by means of mean shift processing. Then, the coordinate assigning unit 310 generates representative vector information regarding the components and coordinates of the representative vector VE ₂ and outputs the representative vector information to the corresponding point selection unit 311.

Next, the corresponding point selection unit 311 deletes a plurality of representative vectors VE ₂ having a maximum frequency less than a predetermined value from among the representative vector information and the smoothed histogram information (that is, cuts off). . Then, the corresponding point selection unit 311 generates corresponding point selection information related to the feature point pair constituting the selected representative vector VE ₂ , that is, the representative point pair, and outputs it to the linear geometric parameter fitting unit 303.

  In step S80, the error calculation unit 315 calculates the error values e1 and e2 shown in the above equations (32) and (33) based on the corresponding point selection information and the parameter information given from the parameter selection unit 305. calculate. Next, the error calculation unit 315 generates error value information regarding the calculated error values e1 and e2, and outputs the error value information to the initial parameter selection unit 316 together with corresponding point selection information. In addition, the error calculation unit 315 selects an error value having a small value, and generates minimum error value information regarding the selected error value. Then, the error calculation unit 315 outputs the minimum error value information and the corresponding point selection information to the weight calculation unit 313.

Next, the initial parameter selection unit 316 determines, based on the error value information, which one of the previous frame homography matrixes HL _n−1 , HR _n−1 and unit matrix I is to be used as an initial parameter. Specifically, the initial parameter selection unit 316 sets the matrix with the smaller error value as the initial parameter homography matrix.

  Next, the initial parameter selection unit 316 generates initial parameter information regarding the selected initial parameter, and outputs it to the linear geometric parameter calculation unit 314. Further, the initial parameter selection unit 316 outputs the corresponding point selection information to the coordinate pair holding unit 317 and the coordinate pair integration unit 318. The initial parameter selection unit 316 outputs reset information to the coordinate pair holding unit 317 when a unit matrix is selected as the initial parameter.

  Next, the coordinate pair integration unit 318 integrates the corresponding point selection information given from the initial parameter selection unit 316 and the coordinate pair holding unit 317. The coordinate pair integration unit 318 outputs the integrated corresponding point selection information to the weight calculation unit 313 and the linear geometric parameter calculation unit 314.

Then, the weight calculator 313 calculates the weights, i.e. the weight w _e time axis constraint indicating how much consideration of the time axis constraint in the calculation of the transformation matrix F. Weight calculator 313 generates a time axis constraint weight information about the weight w _e of determined time axis constraint, and outputs the linear geometric parameter calculation unit 314.

Next, the coordinate conversion unit 319 performs coordinate conversion of each representative point pair based on the information given from the initialization unit 315. Specifically, the coordinate conversion unit 319 performs coordinate conversion of the left representative point with the previous frame homography matrix HL _n-1 or unit matrix I, and converts the right representative point with the previous frame homography matrix HR _n-1 or unit. Coordinate transformation is performed with the matrix I. Next, the coordinate conversion unit 319 normalizes the coordinates and initial parameters of the representative point pair after the coordinate conversion. Next, the coordinate conversion unit 319 generates coordinate conversion information regarding the normalized coordinates of each representative point pair and the normalized homography matrices WL and WR, and outputs the coordinate conversion information to the matrix calculation unit 320.

Then, the matrix calculation unit 320, based on the coordinate transformation information, and calculates the operation matrix M ₂ of the formula (34) described above. The matrix calculation unit 320 generates calculation matrix information regarding the calculated calculation matrix M ₂ and outputs the calculation matrix information to the F matrix calculation unit 321.

Next, the F matrix calculation unit 321 calculates the minimum eigenvector v _min of the operation matrix M ₂ based on the operation matrix information. Further, F matrix calculation unit 321, so that the minimum singular value of the minimum eigenvector _{v min} drops to 0, singular value decomposition minimum eigenvector _{v min} in a matrix of 3 × 3. Then, the F matrix calculation unit 321 calculates the transformation matrix F based on the calculated matrix SVD (v _min ) and the above-described equation (36). Further, F matrix calculation unit 321 by shaking the weight _w value several values of _e, for example, _{^{2 2 × w e, 2 ×}} w e, 2 -1 × w e, 2 -2 × w e, converted The matrix F is calculated. Then, the F matrix calculation unit 321 calculates an error value for each of the calculated conversion matrices F using the above-described equation (32). Then, the F matrix calculation unit 321 selects the one with the smallest error value, generates transformation matrix information regarding the selected transformation matrix F, and outputs the transformation matrix information to the F matrix decomposition unit 322 together with the coordinate transformation information.

Next, the F matrix decomposition unit 322 decomposes the transformation matrix F into homography matrices HL and HR and a matrix F ₀ based on the transformation matrix information. Substantially, the F matrix decomposition unit 322 calculates matrices that make the epipolar lines parallel as the homography matrices HL and HR. Next, the F matrix decomposition unit 322 generates linear correction coefficient information regarding the calculated homography matrices HL and HR, and outputs the linear correction coefficient information to the nonlinear geometric parameter fitting unit 304 and the parameter selection unit 305.

In step S90, the Jacobian matrix calculation unit 323 calculates a feature amount vector information, based on the linear correction coefficient information or non-linear correction coefficient information, the Jacobian matrix J ₂ represented by the aforementioned formula (43). The Jacobian matrix calculation unit 323 generates the Jacobian information about Jacobian matrix J ₂ calculated, and outputs the Hessian matrix calculation section 324 together with the feature vector information and the linear correction coefficient information. Then, the Jacobian matrix calculation unit 323 generates Jacobian matrix information regarding the calculated Jacobian matrix J ₂ and outputs the generated Jacobian matrix information to the Hessian matrix calculation unit 324.

Then, Hessian matrix calculation unit 324, based on the information provided by the Jacobian matrix calculation section 323 calculates the Hessian matrix H ₂ represented by the above Expression (45). Then, Hessian matrix computation unit 324 generates the Hessian information about Hessian H ₂ calculated, and outputs the coefficient update unit 325 together with the information provided by the Jacobian matrix calculation section 323.

  Next, the coefficient updating unit 325 calculates each component of the transformation matrix F as a nonlinear correction coefficient coef (nonlinear) based on the information given from the Hessian matrix calculation unit 324 and the above-described equation (46).

  Next, the coefficient updating unit 325 calculates the homography matrices HL and HR by decomposing the nonlinear correction coefficient coef (nonlinear) by the same process as the process of the F matrix decomposing unit 322 described above. Then, the coefficient updating unit 325 generates nonlinear correction coefficient information regarding the calculated homography matrices HL and HR. Then, the coefficient updating unit 325 determines whether the homography matrices HL and HR have converged to a constant value. If the homography matrices HL and HR have converged to a certain value, the nonlinear correction coefficient information is selected as a parameter. When it is determined that the homography matrices HL and HR have not converged, the nonlinear correction coefficient information is output to the Jacobian matrix calculation unit 323. Thereafter, the process of step S90 is performed again.

  In step S100, the parameter selection unit 305 shown in FIG. 16 performs the above-described processing for each of the linear correction coefficient information given from the linear geometric parameter fitting unit 303 and the nonlinear correction coefficient information given from the nonlinear geometric parameter fitting unit 304. The error value is calculated using the equation (32). The parameter selection unit 305 selects information having a smaller error value as parameter information, and outputs the parameter information to the geometric calibration execution unit 306 and the linear geometric parameter fitting unit 303.

In step S110, the geometric calibration execution unit 306 performs geometric calibration on the color calibrated images V _L-1 and V _R-1 based on the parameter information. Specifically, the geometric calibration execution unit 306 performs coordinate conversion of each coordinate point of the color calibrated image V _L-1 using the homography matrix HL, and converts each coordinate point of the color calibrated image V _R-1. Coordinate transformation is performed using the homography matrix HR. The geometric calibration execution unit 306 outputs the images after geometric calibration to the parallax detection unit 4 shown in FIG. 5 as the calibrated images V _L-2 and V _R-2 , respectively.

[Processing by parallax detection unit]
Next, the procedure of the process performed by the parallax detection unit 4 will be described with reference to the flowchart shown in FIG.

In step S120, the DSAD calculation unit 404 determines the left coordinate point and the horizontal parallax d1 based on the calibrated images V _L-2 and V _R-2 and the reliability map information given from the color calibration unit 2. DSAD (−2) to DSAD (+2) are calculated, DSAD information relating to the calculated DSAD (−2) to DSAD (+2) is generated, and output to the minimum value selection unit 405.

  In step S130, the minimum value selection unit 405 selects the minimum DSAD for all the left coordinate points and the horizontal parallax d1 based on the DSAD information. Then, the minimum value selection unit 405 stores the selected DSAD in each node P (x, d) of the DP map for parallax detection shown in FIG.

On the other hand, the horizontal difference calculation unit 411 obtains the calibrated image V _L-2, the pixels of the calibrated image V _L-2, i.e., to calculate the horizontal difference dx for each left coordinate point, calculated horizontal difference Horizontal difference information relating to dx is generated and output to the filter unit 414.

On the other hand, the vertical difference calculation unit 412 obtains the calibrated image V _L-2, the pixels of the calibrated image V _L-2, i.e., calculates the vertical difference dy per left coordinate point, calculated vertical difference Vertical difference information regarding dy is generated and output to the filter unit 414.

On the other hand, the time difference calculation unit 413 acquires the calibrated image V _L-2 and the calibrated image V ′ _{L-2 of} the previous frame, and the pixel of the calibrated image V _L-2 , that is, the left coordinate. A time difference dt is calculated for each point, and time difference information related to the calculated time difference dt is generated and output to the filter unit 414.

  Next, the filter unit 414 applies a low-pass filter to the horizontal difference information, the vertical difference information, and the time difference information. Thereafter, the filter unit 414 outputs these pieces of information to the motion score calculation unit 415.

  Next, the motion score calculation unit 415 performs static motion determination for each left coordinate point based on the horizontal difference information, the vertical difference information, and the time difference information, and determines a motion score according to the result. The motion score calculation unit 415 outputs motion score information regarding the motion score to the anchor vector construction unit 407. The motion score becomes smaller as the luminance variation of the left coordinate point becomes more intense (that is, the motion of the image becomes more intense).

Next, the local histogram creation unit 416 acquires the parallax map DM _n−1 created in the previous frame. The local histogram creation unit 416 generates a local histogram DH shown in FIG. 41 for each coordinate point of the parallax map (these coordinate points correspond to the left coordinate point). The local histogram creation unit 416 generates local histogram information regarding the local histogram DH and outputs the local histogram information to the anchor vector calculation unit 417.

  Next, the anchor vector calculation unit 417 performs the following processing for each left coordinate point based on the local histogram information. That is, as shown in FIG. 41, the anchor vector calculation unit 417 searches for a horizontal parallax d1, that is, a high-frequency parallax d1 'whose frequency is equal to or higher than a predetermined threshold Th2. The anchor vector calculation unit 417 generates anchor vector information indicating a bonus value obtained by multiplying the motion score by a predetermined value and the high-frequency parallax d1 ', and outputs the anchor vector information to the cost calculation unit 408.

  In step S140, the cost calculation unit 408 illustrated in FIG. 48 updates the value of each node P (x, d) of the disparity detection DP map based on the anchor vector information. That is, the cost calculation unit 408 identifies a node (x, d (= d ′)) corresponding to the high frequency parallax d1 ′ for each left coordinate point, and subtracts the bonus value α2 from the score of this node.

  In step S150, the left-eye image horizontal difference calculation unit 418 performs the same processing as the horizontal difference calculation unit 411 described above. Then, the left-eye image horizontal difference calculation unit 418 outputs the horizontal difference information generated by this processing to the weight calculation unit 420.

On the other hand, the right-eye image horizontal difference calculation unit 419 acquires the calibrated image VR _-2 . The right-eye image horizontal difference calculation unit 419 performs the same process as the above _- described horizontal difference calculation unit 411 on the calibrated image VR _-2 . Then, the right-eye image horizontal difference calculation unit 419 outputs the horizontal difference information generated by this processing to the weight calculation unit 420.

Next, the weight calculator 420 calculates the weight wt _{L of} the left coordinate point and the wt _{R of the} right coordinate point for all coordinate points based on the horizontal difference information. The weight calculation unit 420 generates weight information regarding the calculated weights wt _L and wt _R , and outputs the weight information to the route calculation unit 421. The weights wt _L and wt _{R become} smaller at the edge (outline) of the image.

Next, the route calculation unit 421 calculates an accumulated cost up to each node P (x, d) of the DP map for parallax detection based on the weight information given from the weight calculation unit 420. Specifically, the route calculation unit 421 uses the expression described above for the accumulated cost from the start point to the node P (x, d), with the node (0, 0) as the start point and the node (x _maz , 0) as the end point. It is defined as (68) to (70).

Next, the route calculation unit 421 selects the smallest one of the calculated accumulated costs DFI (x, d) _{0 to} DFI (x, d) ₂ and selects the selected one as the node P (x, d). Is the accumulated cost DFI (x, d). The route calculation unit 421 calculates the accumulated cost DFI (x, d) for all the nodes P (x, d) and stores it in the DP map for parallax detection.

  Next, the backtrack unit 410 calculates the shortest path, that is, the path with the minimum accumulated cost from the start point to the end point by tracing back the path with the minimum accumulated cost from the end point to the start point.

On the other hand, local matching unit 402 generates and outputs a disparity map DM _2, the disparity map information about the parallax map DM ₂ generated disparity map integration unit 403. Further, the local matching unit 402 generates vertical parallax information related to the vertical parallax d2 and outputs the vertical parallax information to the global matching unit 401.

In step S160, the back-track portion 410, the shortest path calculated, based on the vertical parallax information supplied from the local matching unit 402 generates a disparity map DM _1. Backtracking portion 410 generates a disparity map information about the parallax map DM ₁ generated, and outputs the disparity map integration unit 403 shown in FIG. 37.

In step S170, the disparity map integration unit 403 first evaluates the disparity map DM ₁ generated by the global matching unit 401 and the disparity map DM ₂ generated by the local matching unit 402 based on the disparity map information. That is, the disparity map integration unit 403, and evaluated by different evaluation methods parallax map DM ₁ generated by the global matching unit 401, by integrating the results, generates a global matching reliability map EM _G . Similarly, the disparity map integration unit 403, and evaluated by different evaluation methods a disparity map DM ₂ generated by the local matching unit 402, by integrating the results, generates a local matching reliability map EM _L To do.

Then, the disparity map integration unit 403, a global matching reliability map EM _G, by comparing the local matching reliability map EM _L, each left coordinate point, the disparity map DM ₁ and parallax map DM ₂ It is determined which is more reliable. Based on the determination result, the disparity map integration unit 403 generates a parallax detection reliability map EM indicating a parallax map with high reliability for each left pixel block.

The disparity map integration unit 403 generates a disparity map DM based on the generated disparity detection reliability map. Here, the horizontal parallax d1 of each left coordinate point of the disparity map DM indicates the value of the higher reliability of the parallax map DM ₁ and disparity maps DM _2.

  The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

DESCRIPTION OF SYMBOLS 1 Autostereoscopic display apparatus 2 Color calibration part 3 Geometric calibration part 4 Parallax detection part 201 Histogram matching part 202 Linear color parameter fitting part 203 Feature amount matching part 204 Nonlinear color parameter fitting part 205 Evaluation part 206 Color calibration execution part 301 Feature amount matching unit 302 Corresponding point refinement unit 303 Linear geometric parameter fitting unit 304 Nonlinear geometric parameter fitting unit 305 Parameter selection unit 306 Geometric calibration execution unit 401 Global matching unit 402 Local matching unit 403 Disparity map integration unit 404 DSAD calculation Unit 405 minimum value selection unit 406 motion description unit 407 anchor vector construction unit 408 cost Calculator 409 channel establishing unit 410 backtracking portion

Claims (11)

  Among the standard image and the reference image in which the same subject is drawn at different horizontal positions, the standard pixel constituting the standard image and the pixel constituting the reference image are present at the same height position as the standard pixel. A horizontal parallax for comparing a first reference pixel and a second reference pixel having a height position different from that of the first reference pixel and detecting a horizontal parallax of the reference pixel based on a result of the comparison An image processing apparatus including a detection unit.   The horizontal parallax detection unit is configured to evaluate a difference between a feature amount in a reference region including the reference pixel and a feature amount in a first reference region including the first reference pixel; A second evaluation value for evaluating a difference between the feature quantity in the reference area and the feature quantity in the second reference area including the second reference pixel, and calculating the first evaluation value and The image processing apparatus according to claim 1, wherein a horizontal parallax of the reference pixel is detected based on the second evaluation value. A correction execution unit that corrects the base image and the reference image before the horizontal parallax is detected;
A correction reliability calculation unit that calculates the reliability of correction by the correction execution unit,
The image processing apparatus according to claim 2, wherein the horizontal parallax detection unit calculates the first evaluation value and the second evaluation value based on the reliability calculated by the correction reliability calculation unit.   The horizontal parallax detection unit calculates a change amount per unit time of the feature amount of the reference pixel, and calculates the first evaluation value and the second evaluation value based on the calculated change amount. The image processing apparatus according to claim 2.   The horizontal parallax detection unit generates a mosaic image by reducing the resolution of the reference image, and calculates a difference between the feature amount of the reference pixel in the reference image and the feature amount of the reference pixel in the mosaic image. The image processing apparatus according to claim 2, wherein the image processing apparatus calculates the first evaluation value and the second evaluation value based on the calculated difference.   In the horizontal parallax detection unit, the horizontal position of the reference pixel and the horizontal parallax of the reference pixel are components of the node, and the first evaluation value among the first evaluation value and the second evaluation value And generating a DP map in which the evaluation value with the smaller absolute value of the difference evaluated by the second evaluation value is the score of the node, and the feature amount between the reference pixel and pixels in the peripheral region of the reference pixel And the difference between the feature amounts of the reference corresponding pixel existing in the same position as the reference pixel and the pixels in the peripheral region of the reference corresponding pixel among the pixels constituting the reference image, and the score of the node , To calculate a cumulative cost from the start point to the node, to calculate a shortest path that minimizes the cumulative cost from the start point to the end point, based on the calculated shortest path, the reference Pixel horizontal Detecting a difference, the image processing apparatus according to claim 2, wherein.   The image processing apparatus according to claim 1, wherein a difference in height position between the second reference pixel and the first reference pixel is limited within a predetermined range. A second horizontal parallax detection unit that detects a horizontal parallax of the reference pixel by a different detection method from the first horizontal parallax detection unit that is the horizontal parallax detection unit;
The reliability of the horizontal parallax of the reference pixel detected by the first horizontal parallax detection unit and the horizontal parallax of the reference pixel detected by the second horizontal parallax detection unit are respectively calculated, and for each reference pixel The image processing apparatus according to claim 1, further comprising: a disparity map integration unit that generates a disparity map indicating the horizontal parallax having the high reliability.   The disparity map integration unit is configured to determine the reliability of the horizontal parallax of the reference pixel detected by the first horizontal parallax detection unit and the horizontal parallax of the reference pixel detected by the second horizontal parallax detection unit, respectively. The image processing apparatus according to claim 8, wherein the image processing apparatus is calculated by a different evaluation method.   Among the standard image and the reference image in which the same subject is drawn at different horizontal positions, the standard pixel constituting the standard image and the pixel constituting the reference image are present at the same height position as the standard pixel. Comparing a first reference pixel and a second reference pixel having a height position different from that of the first reference pixel, and detecting a horizontal parallax of the reference pixel based on a result of the comparison. An image processing method. On the computer,
Among the standard image and the reference image in which the same subject is drawn at different horizontal positions, the standard pixel constituting the standard image and the pixel constituting the reference image are present at the same height position as the standard pixel. A horizontal parallax for comparing a first reference pixel and a second reference pixel having a height position different from that of the first reference pixel and detecting a horizontal parallax of the reference pixel based on a result of the comparison A program that realizes the detection function.