JP2009059047A - Device, method and program for detecting object - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To enable high-speed processing with a small amount of calculations and detect an object from a digital image at high accuracy. <P>SOLUTION: A device for detecting an object includes: a first detection unit 60 for detecting a plurality of image regions which are likely to be an object of interest from digital image data input; a grouping process unit 70 for extracting, from a plurality of image regions doubly detected for the one object among the plurality of image regions likely to be the one object, the image area that is most likely to be the one object; and a second detection unit 80 for detecting with high accuracy the image region of the object from the image region selected by the grouping process unit 70. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an object detection device, an object detection method, and an object detection program for detecting an object of interest from a digital image.

  Conventionally, in the digital image processing field, various techniques for detecting an object of interest from a digital image have been proposed.

  In these techniques, for example, when detecting a human face as an object of interest, a method using the shape features of each part of the face such as eyes, nose, mouth, or contour, facial tone features, skin color, etc. A method using color characteristics or a method of combining them is used.

  One technique for detecting these objects is a technique called boosting. In this boosting method, a plurality of discriminators with low detection accuracy (weak discriminators) are combined, and each weak discriminator is weighted so as to minimize an error, thereby forming one high-accuracy discriminator. This is a method for detecting an object.

  In the classifier constituted by a large number of weak classifiers, it is possible to detect an object at high speed using learning data generated based on the density characteristics of the object.

  As an example of a technique for detecting an object using such a boosting method, there is a face detector described in Non-Patent Document 1.

  The face detector described in Non-Patent Document 1 is composed of a plurality of weak classifiers, and shows a pattern of light and dark features such as Haar-like features using AdaBoost, which is a boosting technique. By performing learning to extract features that are effective in expressing the brightness of the face and with few errors from rectangular data of multiple shapes and sizes, the face area that is the target object is extracted from the digital image. It is to detect.

  The configuration of the face detector described in Non-Patent Document 1 is shown in FIG.

  As shown in FIG. 5, the face detector 100 divides a plurality of weak classifiers into m classifiers (first classifier 100-1 to m-th classifier 100-m), and from the plurality of weak classifiers. Are configured by cascading m classifiers.

  When the face detector 100 detects an object from a digital image, first, the luminance component (luminance image) of the digital image to be detected is input to create an integral image.

  This integrated image is generated by, for example, performing the work of adding the luminance value of its own pixel to the sum of the luminance values of the pixels one pixel above and one pixel left from the image in order from the lower left of the image. It is an image in which the pixel value at an arbitrary position is the sum of the luminance values of the lower left rectangular area.

  By generating this integrated image, the sum of the luminance values of the rectangular area can be calculated at high speed only by adding and subtracting the luminance values of the pixels at the four corners of the desired rectangular area in the image.

  Next, a window image composed of 24 × 24 pixel images to be searched is sequentially scanned and cut out from the input luminance image.

  Next, this cut-out window image is input to the detector 100, and from the first weak classifier 100-1 (1) to the weak classifier 100-1 (n) in the first first classifier 100-1. The n weak classifiers sequentially determine whether the window image is a face area or a non-face area, and outputs the result according to the reliability of the first classifier 100-1. An evaluation value that is a result of adding the weights is calculated.

  Next, it is determined whether or not the calculated evaluation value is higher than a threshold value set in advance by learning data. If the determination result is high, it is determined that the window image is a face area, and While data is sent to the 2 classifier 100-2, if the window image is lower than the threshold, it is determined that the window image is not a face area, and the determination process is terminated to speed up the process.

  Next, the second classifier 100-2 calculates an evaluation value in the same manner as the first classifier 100-1, and when this evaluation value is higher than the set threshold value, the third classifier 100-3 further stores data. Is sent out.

  The window image is sequentially cut out by scanning the input luminance image vertically and horizontally, it is determined whether or not the window image has a face area, and the window image that has passed to the last m-th classifier 100-m is determined. It is detected as an image of the face area.

  Since the size of the face in the digital image to be detected varies, the digital image is reduced in stages, and the above-described integral image generation and face area detection processing are performed, thereby achieving high accuracy. Perform face detection with.

  The above detection method is also described in Patent Document 1.

  Patent Document 1 describes a facial expression recognition system and a learning method for facial expression recognition, in which a face detection method is described.

  In the learning of the detector for face detection using Adaboost described in Patent Document 1, it is estimated that the performance is higher than others among all weak hypotheses (synonymous with the weak classifier in Non-Patent Document 1). Weak hypotheses are selected, and new weak hypotheses are generated based on statistical properties by the selected weak hypotheses.

  Next, the configuration of the face detector described in Patent Document 1 is shown in FIG.

  In the face detector 200 described in Patent Document 1, a plurality of selected weak hypotheses (first weak hypothesis 200-1 to nth weak hypothesis 200-n) are the same as the weak classifier described in Non-Patent Document 1. However, when the face detector 200 detects an object from a digital image, every time a determination result based on one weak hypothesis (= weak classifier) is output. It is determined whether or not there is a face area in the window image.

  As the determination result by each weak hypothesis, every time the estimated value in each weak hypothesis is calculated, “the estimated value of the discrimination result indicating whether the window image is a face region × the reliability of the weak hypothesis” is added. Calculated as an evaluation value.

  Furthermore, based on the calculated evaluation value and a threshold value for canceling the process calculated in advance by learning, when it is determined that the face area is not clearly a face area, the determination process is terminated to increase the speed. .

  The processing proceeds in this way, and the window image that has passed through to the last weak hypothesis 200-n is detected as an image of the face area.

  In the face detectors described in Non-Patent Document 1 and Patent Document 1, after the discrimination processing is completed for all window images in the digital image, the area detected as the area indicating the face area is in the vicinity. If there are two or more overlapping areas, the area with the smaller evaluation value of the two overlapping areas is considered to be less likely to be a face area, and the area with the larger evaluation value is deleted. Select as an image. Alternatively, an area obtained by averaging them is extracted. This process is repeated until there is no overlap, and the last selected image is detected as a face area image.

  Further, in Patent Document 2, the discriminators are not connected in a line as in Non-Patent Document 1 or Patent Document 1, but as shown in FIG. 7, each node as a discriminator is connected in a network form. 300 is described.

  When detecting an object from a digital image by the detector 300, for example, a path passing through a plurality of nodes such as a path 301 is generated, and a face or the like is detected at each node of the path with respect to the input window image. An accumulation of evaluation values that evaluate whether or not the object is an object is obtained, and an estimation value of an identification error for a path identification result is calculated.

  If the lowest identification error among the identification results in a plurality of paths is smaller than a preset threshold value, the identification process is terminated, and it is determined that there is an object there.

Even when a face region is determined by the above processing, the processing is speeded up by terminating the processing when the identification error becomes smaller than a preset threshold value. If the error does not become smaller than the threshold value, path generation and evaluation are continued, and the process is terminated when the number of paths reaches a preset number.
Paul Viola, Michael Jones, `` Robust Real-Time Face Detection '', International Journal of Computer Vision 57 (2), 137-154, 2004 JP 2005-44330 A JP 2006-350645 A

  However, in the face detector 100 of Non-Patent Document 1, for example, when the input digital image is a luminance image having a width of 320 × 240 pixels in height, the image is reduced to 10 steps by 4/5 times to obtain an integrated image. If a determination process is performed by cutting out a window image having a width of 24 × height of 24 pixels in each reduced image and scanning by skipping one pixel in the vertical and horizontal directions, determination is made on nearly 40,000 window images. Processing will be performed.

  In addition, the learning data used for the determination process is generally formed by hundreds to thousands of feature quantities (6060 in the case of Non-Patent Document 1), and the weak classifier is the maximum for one window image. A discrimination process for the number of pieces of learning data is performed.

  In this discrimination process, when the clipped window image is completely different from the face and shade characteristics, such as a plain wall with low contrast, the first classifier determines that the face is not a face and the process is terminated. In the case of a window image having, it often proceeds to a subsequent classifier, and the amount of calculation for discrimination may become enormous.

  Further, in the vicinity of a window image determined to be a face area, a plurality of window images determined to be a face area by the same detection target are often detected.

  That is, when one face area is detected, discrimination processing is performed for the number of learning data in a plurality of neighboring window images, and the calculation amount is enormous.

  The face detector 200 of Patent Literature 1 has a threshold value for aborting the process for each weak hypothesis (= weak discriminator), but the timing at which the discrimination is aborted is determined by Non-Patent Literature 1. The calculation amount for detecting the object is enormous.

  In the detector 300 of Patent Document 2, a plurality of discriminators are arranged in a network, but the same discriminator (node) arranged at the beginning of the network passes through the same discriminator many times during detection processing. Not efficient.

  In addition, in the shape of the detector 300 of Patent Document 2, since it cannot be cut off in the middle of the node, as a result, it passes through many discriminators in order to determine whether or not it is an object. Significant speedup is not expected.

  Therefore, the present invention has been made in view of the above problems, and is an object detection device and an object that can be processed at high speed with a small amount of calculation and can detect an object from a digital image with high accuracy. An object is to provide an object detection method and an object detection program.

  In order to achieve the above object, an object detection apparatus (1) of the present invention detects an image area including an object from input digital image data, and uses the digital image data as a window of a predetermined size. A plurality of first evaluation values indicating the probability that the object is included for each window are calculated while scanning, and a plurality of candidate image regions are detected based on the calculated first evaluation values. Of the plurality of candidate image areas detected by the detection means (60) and the first detection means (60), at least when there is a candidate image area where the image part overlaps, these overlap. A grouping processing unit (7) that selects a candidate image region having the highest evaluation value from candidate image regions and extracts the selected candidate image region and a candidate image region in which the image portion does not overlap. ) And a plurality of second evaluation values indicating the probability that the object is included for each of the plurality of candidate image areas extracted by the grouping processing means (70), and the calculated second evaluation values And second detection means (80) for detecting the image area based on the value.

  Further, the first detection means (60) of the object detection device (1) according to the present invention uses the first learning data including a plurality of discriminators related to the object by the boosting method. While calculating an evaluation value, the second detection means (80) calculates the second evaluation value by a boosting method using second learning data including a plurality of discriminators related to the object. It may be.

  In addition, the number of the plurality of classifiers in the first learning data of the object detection device (1) of the present invention may be smaller than the number of the plurality of classifiers in the second learning data.

  The object detection method of the present invention is an object detection method for detecting an image region including an object from input digital image data, while scanning the digital image data with a window of a predetermined size, A first detection step (60) of calculating a plurality of first evaluation values indicating the probability of the object being included for each window and detecting a plurality of candidate image regions based on the calculated first evaluation values. ) And a plurality of candidate image areas detected in the first detection step (60), at least when there is a candidate image area where the image part overlaps, the most candidate image area is selected from these overlapping candidate image areas. Grouping processing step of selecting a candidate image region having a high evaluation value and extracting the selected candidate image region and a candidate image region in which the image portion does not overlap 70) and, for each of the plurality of candidate image regions extracted in the grouping processing step (70), a plurality of second evaluation values indicating the probability of the object being included are calculated, and these calculated second values are calculated. And a second detection step (80) for detecting the image area based on the evaluation value.

  Further, in the first detection step (60) of the object detection method of the present invention, the first evaluation value is obtained by a boosting method using first learning data comprising a plurality of discriminators related to the object. In the second detection step (80), the second evaluation value may be calculated by a boosting method using second learning data including a plurality of discriminators related to the object. .

  Further, the number of the plurality of discriminators in the first learning data of the object detection method of the present invention may be smaller than the number of the plurality of discriminators in the second learning data.

  An object detection program of the present invention is an object detection program for causing a computer to execute a process of detecting an image area including an object from input digital image data, wherein the digital image data is While scanning with a window of a predetermined size, a plurality of first evaluation values indicating the probability that the object is included for each window are calculated, and a plurality of candidate image regions are calculated based on the calculated first evaluation values. A first detection step (60) to be detected, and a plurality of candidate image regions detected in the first detection step (60), when there is a candidate image region having at least overlapping image portions. The candidate image region having the highest evaluation value is selected from the overlapping candidate image regions, and the selected candidate image region and the image portion are not overlapped. A grouping process step (70) for extracting candidate image areas, and a plurality of second evaluation values indicating the probability that the object is included for each of the plurality of candidate image areas extracted in the grouping process step (70). And a second detection step (80) for detecting the image area based on the calculated second evaluation value, and causing the computer to execute.

  In the first detection step (60) of the object detection program of the present invention, the first evaluation value is obtained by a boosting method using first learning data composed of a plurality of discriminators related to the object. And the second detection step (80) causes the computer to calculate the second evaluation value by a boosting method using second learning data comprising a plurality of discriminators related to the object. You may make it perform.

  In addition, the number of the plurality of discriminators in the first learning data of the object detection program of the present invention is made smaller than the number of the plurality of discriminators in the second learning data to be executed by the computer. May be.

  According to the object detection device, the object detection method, and the object detection program of the present invention, it is possible to perform high-speed processing with a small amount of calculation and detect an object from a digital image with high accuracy.

  An object detection apparatus according to an embodiment of the present invention will be described in detail with reference to the drawings.

  The object detection apparatus according to the present embodiment uses a conventionally used boosting method, and as shown in FIG. 1, the first detection unit 60 performs high-speed rough detection processing from image data to be detected to detect the object. By detecting a candidate area and extracting a candidate area for the target object by the grouping processing unit 70, the second detection unit 80 performs a highly accurate detection process on the extracted candidate area of the target object. Thus, the region position of the object is detected with high accuracy with a significantly smaller amount of calculation than in the prior art.

  Note that the “rough detection process at high speed” performed by the first detection unit 60 is a trade-off between the speed of the detection process and the detection accuracy, and allows the detection target to be erroneously detected and the detection process speed to be the detection accuracy. It is a priority.

<Configuration of Object Detection Device According to One Embodiment>
A schematic configuration of the object detection apparatus 1 according to the present embodiment will be described with reference to FIG.

  As illustrated in FIG. 2, the object detection device 1 includes an input unit 10, an image scale change unit 20, an integral image generation unit 30, a processing data storage unit 40, a learning data storage unit 50, and a first detection. Unit 60, grouping processing unit 70, second detection unit 80, and output unit 90.

  The input unit 10 inputs image data composed of luminance components of search target digital image data.

  The image scale changing unit 20 changes the scale of the input image data.

  The integrated image generation unit 30 generates integrated image data of each image data whose scale has been changed.

  The processing data storage unit 40 includes integrated image data generated by the integrated image generation unit 30, target object window image information calculated by a first detection unit 60 described later, and target objects extracted by the grouping processing unit 70. The candidate window image information and the object area window image information calculated by the second detection unit 80 are stored.

  The learning data storage unit 50 includes a plurality of feature amounts indicating features of the target object, their sizes and positions, reliability (weights) of each feature amount, and search target images generated by the prior learning. Learning data such as a threshold value of an evaluation value (to be described later) for determining whether or not it is determined is stored. The feature amount here is, for example, a black-and-white binary pattern showing a local lightness difference such as a Haar-like feature, or a shading pattern at an arbitrary position of two or more points on the window.

  The 1st detection part 60 has the 1st scanning part 61 and the 1st discrimination | determination part 62, and detects the candidate of a target object area | region by a boosting method.

  The first scanning unit 61 scans the input luminance component image data, and sequentially cuts out a window image, which is a rectangular image to be searched, while shifting the window image vertically and horizontally by predetermined pixels.

  The first discriminating unit 62 includes a plurality of discriminators (not shown) provided for each feature amount of the learning data stored in the learning data storage unit 50, and each discriminator includes a first scanning unit 61. For the sequentially extracted window images, the gradient of the shade corresponding to the feature amount is calculated using the integrated image stored in the processing data storage unit 40, and the object in the classifier is calculated from the gradient of the shade. Evaluate materiality and output an evaluation value α. Further, by weighting the evaluation value α according to the reliability of each discriminator, an evaluation value β indicating the degree of objectness is calculated, and a value obtained by accumulating the evaluation value β calculated by each discriminator is calculated. The first evaluation value is calculated for each window image.

  And the 1st discrimination | determination part 62 compares the calculated 1st evaluation value and the threshold value of the evaluation value memorize | stored in the learning data storage part 50, and if the 1st evaluation value exceeds this threshold value, the said window image Is detected as an object candidate window image that is a candidate for the object region by determining that “it may be an object”, and the position information and the first evaluation value of this window image in the image data being searched Are sent to the processing data storage unit 40 as object candidate window image information.

  The grouping processing unit 70 includes an overlapping window image detection unit 71 and a candidate window image selection unit 72.

  The overlapping window image detection unit 71 detects an object candidate window image at a nearby position that is detected redundantly by the first determination unit 62 for one object region in the luminance component image data.

  The candidate window image selection unit 72 detects a plurality of object candidate window images at neighboring positions detected in the overlapped window image detection unit 71 by the first determination unit 62 overlapping with one object region. The object candidate window image having the highest evaluation value is selected from the object candidate window image group, and the object candidate window image information of the object candidate window image is sent to the processing data storage unit 40. The object candidate window image information of the candidate window image information that has not been selected and is stored in the processing data storage unit 40 is deleted. Note that the candidate window image selection unit 72 may calculate an average position from the position of the target object candidate window image group and send it to the processing data storage unit 40.

  The second detection unit 80 includes a second scanning unit 81 and a second determination unit 82, and detects a candidate for the object region by a boosting method.

  The second scanning unit 81 acquires the position information of the object candidate window image extracted by the grouping processing unit 70 from the processing data storage unit 40, and based on the acquired position information, the object candidate window image is obtained from the luminance component image data. Cut out.

  The second discriminating unit 82 is provided for each feature amount of the learning data stored in the learning data storage unit 50 and has a sufficient number of discriminators (FIG. 5) having sufficient reliability to detect the object with high accuracy. In each discriminator, the integrated image stored in the processing data storage unit 40 is the gradient of the area corresponding to the feature amount of the window image cut out by the second scanning unit 81. Is used to evaluate the object-likeness of the discriminator from the shade gradient, and an evaluation value γ is output. Further, by adding a weight according to the reliability of each discriminator to the evaluation value γ, an evaluation value λ indicating the degree of objectness is obtained, and a value obtained by accumulating the evaluation value λ obtained by each discriminator For each window image as a second evaluation value.

  And the 2nd discrimination | determination part 82 compares the calculated | required 2nd evaluation value and the threshold value of the evaluation value memorize | stored in the learning data memory | storage part 50, and if the 2nd evaluation value exceeds this threshold value, the said window image Is determined to be “object” and detected as an object area window image, and the position information and the second evaluation value of the object area window image in the input luminance component image data are used as the object area window image. The information is sent to the processing data storage unit 40 as information.

  The number of discriminators constituting the second discriminating unit 82 is appropriately changed according to the purpose of use, but is configured with a number of discriminators having sufficient reliability for detecting an object with high accuracy.

  Note that the first determination unit 62 described above is intended to determine whether or not there is a possibility of being an object, and therefore the height accuracy necessary for final detection is not necessary. Specifically, the average calculation amount per window image can be expressed as “average calculation amount = average number of feature amounts to be used × average calculation amount per one feature amount”. It is sufficient that the average calculation amount is sufficiently smaller than that of the second determination unit 82. Therefore, the first discriminating unit 62 only needs to have a small amount of calculation even if the number of feature amounts is large.

  The output unit 90 reads and outputs the object area window image information stored in the processing data storage unit 40.

<Operation of Object Detection Device According to One Embodiment>
An operation when a face area is detected as an object by the object detection apparatus 1 according to the present embodiment will be described with reference to a flowchart of FIG. 3 and a screen display diagram of FIG.

  First, when the image data a of the luminance component of the digital image data to be searched is input to the input unit 10 of the object detection device 1 (S1), the image data a is set in advance in the image scale changing unit 20. The scale is reduced to p times (S2).

In the present embodiment, the input image data a is reduced by a factor of p as defined by the following equation (1).

(However, q is
0 <q <1
Where r is a constant of
0 ≦ r ≦ T-1 (T: the number of scale change repetitions is a natural number)
It is a natural number that increases by 1 in the range. )
The number of repetitions T of the scale change is set in advance according to the range of the face size to be detected.

  The first time is r = 0, and the input image data a is processed with the same magnification.

  Next, in the integrated image generation unit 30, integrated image data of the image data a is generated and stored in the processing data storage unit 40 (S3).

  Next, in the first scanning unit 61 of the first detection unit 60, a rectangular window having a preset size is used, and a window image that is a rectangular image to be searched is scanned and cut out from the image data a.

  In this embodiment, the size of the rectangular window is 24 × 24 pixels high, and in the first scan, for example, the lower left pixel (X coordinate = 0 of the image data = 0) set as the initial position in the image data to be searched. , Y coordinate = 0) is cut out in a state where the lower left pixel of the window image is set (S4, S5).

  Next, in the first discriminating unit 62, feature amounts are extracted from the window image cut out by the first scanning unit 61 using, for example, the boosting method described in Non-Patent Document 1 and Patent Document 1 described above. The

  In each discriminator of the first discriminating unit 62, the gradient of the shade of the region corresponding to the feature amount of the window image sequentially cut out by the first scanning unit 61 is integrated in the processing data storage unit 40. It calculates using the image, evaluates the object likeness in the discriminator from the gradient of the shade, and outputs the evaluation value α. Further, by weighting the evaluation value α according to the reliability of each discriminator, an evaluation value β indicating the degree of objectness is calculated, and a value obtained by accumulating the evaluation value β calculated by each discriminator is calculated. The first evaluation value is calculated for each window image. Then, in the first discriminating unit 62, a weight corresponding to the reliability of each discriminator is added to the calculated evaluation value α, and an evaluation value β indicating the degree of objectness is calculated (S6).

  Further, in the first discriminating unit 62, a first evaluation value in which the evaluation values calculated by the respective discriminators are accumulated is calculated for each window image, and the first evaluation value is stored in the learning data storage unit 50. It is determined whether a value threshold is exceeded.

  As a result of the determination, if the first evaluation value exceeds this threshold value, the window image is determined as “possibly a face” and is detected as a face candidate window image as a face area candidate (“S7” YES ”), the position information of the window image in the image data being searched and the first evaluation value are stored in the processing data storage unit 40 as face candidate window image information (S8).

  The number of discriminators constituting the first discriminating unit 62 is appropriately changed according to the purpose of use and detection accuracy. In order to detect an object with high accuracy like the number of discriminators in the conventional boosting method. It is composed of a number that is much smaller than the number with sufficient reliability, and the standard of whether or not it may be an object with a smaller amount of calculation than these conventional detectors (allowing false detection) ) Detection processing is performed.

  When the process of step S8 is completed, the process returns to step S5, the rectangular window for cutting out the window image is scanned by shifting by m pixels in the X direction in the image data a, and the window image is cut out, and the processes of steps S6 to S8 are performed. Is done. The repetitive processing from step S5 to step S8 is performed j times to the right end of the search target image data a (S9, S10).

  In addition, the repetition process of steps S9 and S10 is repeated i times up to the upper end of the image data a to be searched by scanning by shifting n pixels at a time in the Y direction in the image data a (m, n: An integer greater than or equal to 1) (S11, S12).

  When the search process of the face candidate window image in the search target image data a is completed by the processing in steps S5 to S12, the process returns to step S2 (“NO” in S13), and the value of r in the above equation (1) is added. R = 1, the image data a is reduced based on the value of the constant q, and the processes of steps S3 to S12 are repeated T times set as the number of scale change repetitions.

  When the detection process of the face candidate window image from the reduced image data in the first detection unit 60 is repeated T times (“YES” in S13), the process is stored in the processing data storage unit 40 by the process of the first detection unit 60. The candidate face window image information is read out by the overlapping window image detecting unit 71 of the grouping processing unit 70, and is detected by overlapping with respect to one face candidate area in each image data. Is detected.

  Here, whether or not the window images overlap with each other in one image candidate area in each image data is determined by whether or not the window images to be compared are within a predetermined distance. .

  When the face candidate window image group detected in duplicate is detected (“YES” in S14), the evaluation value of each face candidate window image group is processed in the candidate window image selection unit 72 by the processing data storage unit 40. The face candidate window image having a high evaluation value is selected from the detected face candidate window image group and the face candidate window image information of the face candidate window image having a low evaluation value is processed data storage unit 40. Is deleted and discarded (S15).

  The processes of step S14 and step S15 are repeated until only one face candidate window image is selected for one face candidate area in each image data.

  When only one face candidate window image is selected for one face candidate region in each image data from the group of face candidate window images detected in duplicate (“NO” in S14), the selected face Face candidate window image information of the candidate window image is stored in the processing data storage unit 40 (S16).

  FIG. 4A shows an example when the extracted face candidate window image after grouping is displayed as a rectangular region with a thick frame on the image data to be searched. In FIG. 9A, since the face determination area detection process is performed with low accuracy in the first determination unit 62, a plurality of areas including the actual face area and the area other than the face are selected as the face candidate areas. Is shown.

  Next, the face candidate window image information stored in step S16 in the second scanning unit 81 of the second detection unit 80 is read from the processing data storage unit 40 (S17), and the position information is associated with the image data a. The face candidate window image to be cut out is cut out (S18).

  Next, in the second discriminating unit 82, the evaluation value is calculated by each discriminator and weighted for the window image cut out by the second scanning unit 81 as in step S6 (S19).

  In each discriminator of the second discriminating unit 82, the integrated image stored in the processing data storage unit 40 is the gradient of the shade corresponding to the feature amount of the window image cut out by the second scanning unit 81. Is used to evaluate the object-likeness of the discriminator from the shade gradient, and an evaluation value γ is output. Further, by adding a weight according to the reliability of each discriminator to the evaluation value γ, an evaluation value λ indicating the degree of objectness is obtained, and a value obtained by accumulating the evaluation value λ obtained by each discriminator For each window image as a second evaluation value.

  In the second discriminating unit 82, as in the processing in the first discriminating unit 62, a second evaluation value obtained by accumulating the evaluation values calculated by the respective discriminators is calculated for each window image, and the second evaluation value is learned. It is determined whether the threshold value of the evaluation value stored in the data storage unit 50 is exceeded (S20).

  As a result of the determination, if the second evaluation value exceeds this threshold value, the window image is determined to be “face” and detected as a face area window image (“YES” in S20), and this window image image a The position information of the window image and the second evaluation value are stored in the processing data storage unit 40 as face area window image information (S21).

  The processes in steps S17 to S21 are performed for the face candidate window images corresponding to all the position information stored in step S16 (“YES” in S22), and the process for the face candidate window images corresponding to all the position information is performed. Is completed (“NO” in S22), the position information of the face area window image information stored in step S21 is read by the output unit 90 and output to the outside (S23).

  FIG. 4B shows an example when the position information of the output face area window image information is displayed as a thick rectangular area on the image data to be searched.

  Here, in the second discriminating unit 82, the face candidate area detection process is performed with high accuracy from the face candidate window image selected in FIG. 4A, so that the actual face area is selected. It is shown.

  According to the present embodiment described above, the first detection unit detects an area that may be a target object (allowing a loose reference) from image data with a small amount of calculation, and further performs grouping. The processing unit narrows down the window image to be detected by deleting the target object window image that is detected redundantly with respect to one target region in the image data, and then the second detection unit has high accuracy. Therefore, the object can be detected at a high speed with a remarkably small amount of calculation compared to the detection processing by the conventional boosting method, and the object can be detected from the digital image with high accuracy.

  In the present embodiment, the learning data used in the first determination unit 62 and the second determination unit 82 and the threshold value are described as being the same value, but the first determination unit 62 and the second determination unit 82 It is desirable that the first and second learning data and the first and second threshold values can be set for each.

  Then, as the first and second learning data, the first several discriminators of the learning data having a series of sufficiently accurate discriminators are used as the first learning data, and the discriminators from then on until the last are discriminated. 2 learning data, the first and second learning data may be the same feature amount, the first learning data uses a feature amount of a shape that is easy to specify a position that is likely to be a face, for example, In the case of a rectangular feature, learning may be performed using different feature amounts, such as using a feature amount that requires less calculation amount with a low frequency.

  Furthermore, in the present embodiment, the first discriminating unit 62 and the second discriminating unit 82 use all the discriminators to obtain the first evaluation value and the second evaluation value, and then compare with the threshold value. However, the present invention is not limited to this, and the accumulated evaluation value up to the point where it passed through the classifier in the middle is retained as threshold learning data corresponding to the middle stage. If the discrimination process for the window image is discontinued, the speed can be further increased.

  Moreover, the computer can be made to function as a target object detection apparatus by programming the function structure of the target object detection apparatus of this embodiment as a target object detection program, and incorporating it into a computer.

It is a block diagram which shows the structure of the target object detection apparatus 1 by one Embodiment of this invention. It is a block diagram which shows schematic structure of the target object detection apparatus 1 by one Embodiment of this invention. It is a flowchart which shows operation | movement of the target object detection apparatus 1 by one Embodiment of this invention. The screen display figure (a) which shows the state which displayed the candidate object area | region detected by the 1st detection part 60 of the target object detection apparatus 1 by one Embodiment of this invention on the image data of a detection target, and a target object It is a screen display figure (b) which shows the state where the object field detected by the 2nd detection part 80 of detecting device 1 was displayed on the image data of a candidate for detection. It is a block diagram which shows the structure of the conventional face detector described in the nonpatent literature 1. It is a block diagram which shows the structure of the conventional face detector described in patent document 1. FIG. It is a block diagram which shows the structure of the conventional face detector described in patent document 2.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Object detection apparatus 10 ... Input part 20 ... Image scale change part 30 ... Integrated image generation part 40 ... Processing data storage part 50 ... Learning data storage part 60 ... 1st detection part 61 ... 1st scanning part 62 ... 1st Discriminator 70 ... Grouping processor 71 ... Duplicate window image detector 72 ... Candidate window image selector 80 ... Second detector 81 ... Second scanning unit 82 ... Second discriminator 90 ... Output unit

Claims (9)

In the object detection device for detecting an image area including the object from the input digital image data,
While scanning the digital image data with a window of a predetermined size, a plurality of first evaluation values indicating the probability that the object is included for each window are calculated, and a plurality of values are calculated based on the calculated first evaluation values. First detection means for detecting a candidate image region of
Among the plurality of candidate image areas detected by the first detection means, if there is a candidate image area where at least the image portion overlaps, the candidate having the highest evaluation value from these overlapping candidate image areas Grouping processing means for selecting an image area and extracting the selected candidate image area and a candidate image area in which the image portion does not overlap;
For each of a plurality of candidate image regions extracted by the grouping processing means, a plurality of second evaluation values indicating the probability of the object being included are calculated, and the image regions are based on the calculated second evaluation values. Second detecting means for detecting
An object detection apparatus comprising:   The first detection means calculates the first evaluation value by a boosting method using first learning data comprising a plurality of discriminators related to the object, and the second detection means The object detection apparatus according to claim 1, wherein the second evaluation value is calculated by a boosting method using second learning data including a plurality of classifiers related to the object.   The object detection apparatus according to claim 2, wherein the number of the plurality of classifiers in the first learning data is smaller than the number of the plurality of classifiers in the second learning data. In an object detection method for detecting an image region including an object from input digital image data,
While scanning the digital image data with a window of a predetermined size, a plurality of first evaluation values indicating the probability that the object is included for each window are calculated, and a plurality of values are calculated based on the calculated first evaluation values. A first detection step of detecting a candidate image region of
Among the plurality of candidate image areas detected in the first detection step, if there is a candidate image area where at least an image portion overlaps, the candidate having the highest evaluation value from these overlapping candidate image areas A grouping process step of selecting an image region and extracting the selected candidate image region and a candidate image region in which the image portion does not overlap;
For each of the plurality of candidate image areas extracted in the grouping processing step, a plurality of second evaluation values indicating the probability of the object being included are calculated, and the image areas are based on the calculated second evaluation values. A second detection step of detecting
The object detection method characterized by having.   In the first detection step, the first evaluation value is calculated by a boosting method using first learning data including a plurality of discriminators related to the object, and the second detection step includes the object 5. The object detection method according to claim 4, wherein the second evaluation value is calculated by a boosting method using second learning data comprising a plurality of classifiers related to an object.   6. The object detection method according to claim 5, wherein the number of the plurality of classifiers in the first learning data is smaller than the number of the plurality of classifiers in the second learning data. An object detection program for causing a computer to execute processing for detecting an image area including an object from input digital image data,
While scanning the digital image data with a window of a predetermined size, a plurality of first evaluation values indicating the probability that the object is included for each window are calculated, and a plurality of values are calculated based on the calculated first evaluation values. A first detection step of detecting a candidate image region of
Among the plurality of candidate image areas detected in the first detection step, if there is a candidate image area where at least an image portion overlaps, the candidate having the highest evaluation value from these overlapping candidate image areas A grouping process step of selecting an image region and extracting the selected candidate image region and a candidate image region in which the image portion does not overlap;
For each of the plurality of candidate image areas extracted in the grouping processing step, a plurality of second evaluation values indicating the probability of the object being included are calculated, and the image areas are based on the calculated second evaluation values. A second detection step of detecting
An object detection program for causing the computer to execute.   In the first detection step, the first evaluation value is calculated by a boosting method using first learning data including a plurality of discriminators related to the object, and the second detection step includes the object The object detection program according to claim 7, wherein the computer is executed to calculate the second evaluation value by a boosting method using second learning data including a plurality of classifiers related to an object.   9. The object detection program according to claim 8, wherein the number of the plurality of classifiers in the first learning data is made smaller than the number of the plurality of classifiers in the second learning data to be executed by the computer. .

Images