JP5335554B2 - Image processing apparatus and image processing method - Google Patents

Description

  The present invention relates to an image processing apparatus and an image processing method suitable for detecting a specific subject or a part of a subject from still images and moving images in digital image equipment and image processing software.

  A technique for automatically detecting a specific subject pattern from an image can be applied to various fields such as image search, object detection, object recognition, and object tracking. An example of such a technique is proposed in Non-Patent Document 1. In this technique, a rectangular small area (hereinafter referred to as a detection window) is first extracted from an input image, and it is determined whether or not a face is included in the detection window. In this determination, a detection window is passed through a detector configured by connecting a plurality of strong discriminators in a cascade type. If it is determined that all the strong classifiers are faces, it outputs that there is a face in the detection window, and otherwise outputs that there is no face in the detection window. Each strong classifier includes a plurality of weak classifiers.

  However, in order to actually detect a face or a person with high accuracy using the conventional technique described in Non-Patent Document 1, hundreds to thousands of weak classifiers are required. The number of discriminating processes will be very large. As a result, the processing time becomes longer.

  On the other hand, in the extraction (extraction) of the detection window input to the detector, first, the input image is resized to a reference image, and then the reference image is converted into a luminance image or the like. Then, reduced images of several stages are generated with respect to the size of the luminance image or the like, and the detection window is cut out by raster scanning so that the detection window scans all regions in these reduced images. Therefore, the number of detection windows input to the detector is enormous.

  The reason why the reduced image is generated by resizing the reference image to several sizes and the detection window is raster-scanned for the reduced image is as follows. That is, the size of the subject in the reference image varies, but the detection window to be cut out has a specific size. For this reason, when the subject in the reference image is an area larger than the detection window size, the subject cannot be detected. Therefore, a reduced image is generated by reducing the size of the reference image by several steps, and subjects of various sizes are detected by raster scanning the detection window for all the reduced images.

  As described above, in the technique described in Non-Patent Document 1, a huge number of detection windows cut out from an input image need not be subjected to discrimination processing in a detector composed of hundreds to thousands of weak classifiers. In other words, the processing cost (processing time) is very high.

  For such a technique, a method of shortening the processing time by simply performing raster scanning of the detection window by pixel thinning is conceivable. However, in this method, a position where a characteristic pattern is supposed to be determined may be skipped. As a result, detection omissions increase and detection accuracy decreases.

  In order to deal with these problems, Patent Document 1 describes a technique for learning a subject pattern whose position is shifted in addition to normal subject pattern learning, instead of performing pixel thinning in the raster scan of the detection window. Yes. In this technique, all these learning results are held as dictionary data.

  However, there is a possibility that the positional deviation in the case of pixel thinning may be shifted in at least four directions, up, down, left, and right. Therefore, learning for each positional deviation must be performed. As a result, the dictionary size becomes larger than that of a normal dictionary, which causes another problem.

JP 2007-58722 A

P. Viola and M. Jones, "Robust Real-time Object Detection", SECOND INTERNATIONAL WORKSHOP ON STATISTICAL AND COMPUTATIONAL THEORIES OF VISION, July 13 2001

  An object of the present invention is to provide an image processing apparatus, an image processing method, and the like that can detect a subject with high accuracy and in a short time while suppressing an increase in dictionary size.

  As a result of intensive studies to solve the above problems, the present inventor has come up with various aspects of the invention described below.

An image processing apparatus according to the present invention includes a dictionary storage unit that stores dictionary information including position and shape information of a local region for determining a predetermined subject, and a position of a detection window in the image based on the dictionary information. Position control means for controlling, determination means for determining a local area in the detection window based on the dictionary information, local feature quantity calculation means for calculating a feature quantity in the determined local area, and the feature quantity and have a, a determination unit configured to determine whether the predetermined object is included in the detection window based on said dictionary information, said determination means, the local region width is wide enough horizontal scanning The position of the local region for which the feature amount is to be calculated is determined so that the interval becomes wider, or the vertical scanning interval becomes wider as the vertical width of the local region becomes wider. And determining the position of the local region for which to calculate the amount.

An image processing method according to the present invention controls the position of a detection window in an image based on dictionary information stored in a dictionary storage means and including information on the position and shape of a local region for determining a predetermined subject. A position control step; a determination step of determining a local region in the detection window based on the dictionary information; a local feature amount calculation step of calculating a feature amount in the determined local region; the feature amount and the dictionary It has a, and determining whether the predetermined object is included in the detection window based on the information, in the determining step, the scanning interval of approximately the width is wider horizontal direction of the local region is wide The position of the local region that is the target for calculating the feature amount is determined so that the vertical scanning interval becomes wider as the vertical width of the local region is larger. And determining the position of the local region for which to calculate the feature quantity.

  According to the present invention, when the presence / absence of the subject is determined by scanning the detection window, the local region for which the feature amount is calculated is determined according to the shape, so that the determination can be performed with high accuracy and in a short time. It can be carried out.

1 is a block diagram illustrating a configuration of an image processing apparatus according to a first embodiment of the present invention. It is a figure which shows the structure of the cascade type detector used in 1st Embodiment. It is a figure which shows the structure of the strong discriminator used by 1st Embodiment. It is a figure which shows the process which extracts the detection window (partial area | region) of a predetermined | prescribed magnitude | size from a reduced image. 3 is a flowchart illustrating an operation of the image processing apparatus according to the first embodiment. It is a flowchart which shows the content of the person determination process of step S505. It is a figure which shows the specific example of the relationship between a strong discriminator and a weak discriminator. It is a figure which shows an example of a data structure. It is a figure which shows the basic principle of the object likelihood calculation process in embodiment of this invention. It is a figure which shows the basic principle of the object likelihood calculation process in embodiment of this invention. It is a flowchart which shows the content of the object likelihood calculation process of step S603 in 1st Embodiment. It is a flowchart which shows the content of the local feature-value calculation process of step S1104. It is a figure which shows the relationship between the gradient intensity | strength and gradient direction of the luminance value regarding the pixel (u, v) to which attention is paid. It is a figure which shows the example of a gradient direction histogram. It is a figure which shows the example of a block area. It is a flowchart which shows the content of the object likelihood calculation process of step S603 in 2nd Embodiment. It is a figure which shows the example of a conversion table in case a detection window is 60x120.

  Hereinafter, embodiments of the present invention will be specifically described with reference to the accompanying drawings.

(First embodiment)
First, a first embodiment of the present invention will be described. FIG. 1 is a block diagram showing a configuration of an image processing apparatus according to the first embodiment of the present invention.

  As shown in FIG. 1, the image processing apparatus includes an image input unit 101, an image storage unit 102, a reduced image generation unit 103, a reduced image setting unit 104, a local feature amount calculation unit 105, a position control unit 106, a dictionary storage, and the like. A unit 107, a determination unit 108, and a determination result storage unit 109 are provided. Furthermore, a control unit (not shown) for controlling these operations is also provided.

  The image input unit 101 inputs an image to be detected. The image storage unit 102 stores the image (input image) input by the image input unit 101. The reduced image generation unit 103 generates reduced images of several stages from the input image. The reduced image setting unit 104 sets one reduced image from several stages of reduced images. The dictionary storage unit 107 stores dictionary data (dictionary information) learned in advance. The position control unit 106 controls the position of the detection window in the reduced image based on the dictionary data. The local feature amount calculation unit 105 calculates a feature amount (local feature amount) of a local region in the detection window. The determination unit 108 determines whether there is a person in the detection window based on the local feature amount. The determination result storage unit 109 stores the position of the person that is the output result of the determination unit 108.

  Next, an operation for detecting the position of a person as a subject from an input image in the image processing apparatus configured as described above will be described. FIG. 5 is a flowchart showing the operation of the image processing apparatus according to the first embodiment.

  First, in step S501, an image is input from the image input unit 101, and the image storage unit 102 reads it. Here, the read image data is, for example, two-dimensional array data composed of 8-bit pixels, and is composed of R, G, B, and three surfaces.

  In step S502, the reduced image generation unit 103 generates image data obtained by reducing the image data to a predetermined magnification. This is because in the present embodiment, detection is sequentially performed on image data of a plurality of sizes in order to cope with detection of persons of various sizes. For example, reduction processing to a plurality of images with different magnifications by about 1.2 is sequentially applied for subsequent detection processing.

  Thereafter, in step S503, the reduced image setting unit 104 sets one of a plurality of reduced images generated by the reduced image generation unit 103.

  Subsequently, in step S504, the position control unit 106 extracts a detection window (partial region) having a predetermined size from the reduced image. Here, this extraction process will be described with reference to FIG. First, the input image 401 is resized to a reference image 402, and then the reference image 402 is converted into an image of a predetermined format, for example, a luminance image 403, used in a later determination process (step S505). Then, a reduced image 404 of several stages is generated with respect to the size of the luminance image 403, and the detection window (partial) is raster-scanned so that the detection window scans all areas in the reduced image 404. Area) 405 is cut out. Therefore, when a person is identified by cutting out a detection window from an image with a large reduction ratio, a large person is detected in the image.

  In step S505, the local feature amount calculation unit 105 and the determination unit 108 determine whether a person is included in the detection window. Details of this determination processing will be described later.

  Thereafter, in step S506, the control unit determines whether or not the detection window has scanned all positions in the image. If all positions have been scanned, the process proceeds to step S507. If not, the process proceeds to step S504, and the processes from step S504 to step S506 are repeated until the scanning of all positions is completed.

  In step S507, it is determined whether or not processing has been completed for all of the plurality of reduced images generated by the reduced image generation unit 103 in step S502. If all the processes have been completed, the process proceeds to step S508. If not, the process proceeds to step S503, and the process from step S503 to step S507 is repeated for the next reduced image.

  In step S508, the control unit outputs the detected position of the person to the determination result storage unit 109, and the determination result storage unit 109 stores this.

[Person Determination Processing (Step S505)]
Next, details of the person determination process in step S505 will be described. In this determination process, the cascade detector shown in FIG. 2 is used. FIG. 2 is a diagram showing the configuration of the cascade detector used in the first embodiment. As shown in FIG. 2, this cascade detector is configured by cascade-connecting N strong discriminators 20-1 to 20-N. Each strong discriminator has the configuration shown in FIG. FIG. 3 is a diagram showing the configuration of the strong discriminator used in the first embodiment. The strong classifier includes M weak classifiers 30-1 to 30-M, an adder 311 and a threshold processing unit 312 as shown in FIG. The weak classifiers 30-1 to 30-M output 0 or 1. The adder 311 adds together values obtained by multiplying the signals output from the weak classifiers 30-1 to 30-M by weights set in advance for the weak classifiers 30-1 to 30-M, and outputs the result. To do. The threshold processing unit 312 outputs a discrimination result based on the value output from the adder 311 and a preset threshold. Note that the number M of weak classifiers included in one strong classifier need not be uniform, and the number M of weak classifiers may be different for each strong classifier.

  In step S505, the following processing is executed. FIG. 6 is a flowchart showing the content of the person determination process in step S505.

  First, in step S601, the local feature quantity calculation unit 105 initializes the strong classifier number n to n = 1. The number n is a natural number between 1 and N.

  Next, in step S602, the local feature quantity calculation unit 105 initializes the weak classifier number t in the n-th strong classifier as t = 1, and also adds the subject likelihood of each weak classifier to the sum. A variable Sn for substitution is initialized as Sn = 0. The number t is a natural number between 1 and M.

  Next, in step S603, the local feature amount calculation unit 105 calculates the subject likelihood Lnt (x, y) of the t-th weak discriminator hnt (x, y) in the n-th strong discriminator. Here, the relationship between the strong classifier and the weak classifier will be specifically described. FIG. 7 is a diagram illustrating a specific example of the relationship between the strong classifier and the weak classifier. In this specific example, it is assumed that the strong classifier 1 includes three weak classifiers h11 to h13, and the strong classifier 2 includes four weak classifiers h21 to h24. Further, as shown in FIG. 7, local areas of various positions, sizes, and shapes are set in the detection window, and the weak classifier calculates subject likelihoods in these local areas. Hereinafter, the object likelihoods calculated by the weak classifiers h11 to h13 and h21 to h24 are represented as L11 to L13 and L21 to L24. Details of the subject likelihood calculation process will be described later. The strong discriminators 1 and 2, weak discriminators h11 to h13 and h21 to h24, and information associated therewith are learned in advance and stored in the dictionary storage unit 107 as a data structure as shown in FIG. deep. This data is composed of the number of strong discriminators 801 and the data 802 of each strong discriminator. The data 802 of each strong discriminator includes the number of weak discriminators 803, the data 804 of each weak discriminator, the threshold value 805 for discrimination. Consists of. The data 804 of each weak classifier includes local area information and likelihood conversion LUT (lookup table) as indicated by 806, and the local area information is represented by X at the upper left of the area as indicated by 807. It consists of coordinates, Y coordinate, width and height.

  Note that the shape of the local area corresponding to the weak discriminator need not be a rectangular area. For example, as shown in FIG. 7, a local area 701 or 702 composed of a combination of a plurality of rectangular areas may be used. Further, the number of weak classifiers included in the strong classifier is not limited and may be one or two or more.

  After step S603, in step S604, the local feature amount calculation unit 105 uses the adder 311 to add the subject likelihood Lnt (x, y) acquired in step S603 to the total value Sn.

  Next, in step S605, the weak feature classifier hnt (x, y) that the local feature quantity calculation unit 105 is currently focusing on is the last weak classifier (Mth weak classifier) in the nth strong classifier. ). If it is not the last weak classifier, the process proceeds to step S603, and the process up to step S605 is repeated.

  In the case of the last weak classifier, the process proceeds to step S606, and the determination unit 108 compares the value Sn and the threshold value Tn of the nth strong classifier using the threshold processing unit 312. The threshold value Tn may be obtained in advance by a learning process, stored in the dictionary storage unit 107, and referred to. In the example shown in FIG. 7, the sum S1 of the strong discriminator 1 is S1 = L11 + L12 + L13, and this sum S1 is compared with the threshold T1 of the strong discriminator 1. Further, the total value S2 of the strong discriminator 2 is S2 = L21 + L22 + L23 + L24, and this sum S2 is compared with the threshold value T2 of the strong discriminator 1.

  Then, when Sn <Tn, the determination unit 108 determines that no person is included in the currently focused detection window, ends the processing illustrated in the flowchart of FIG. 6, and returns to the flowchart of FIG. 5. Flow proceeds to step S505. On the other hand, when Sn ≧ Tn (Sn> = Tn), the determination unit 108 determines that the currently detected detection window is a pattern that satisfies the conditions for the determination processing of the nth strong classifier, Flow proceeds to S607.

  In step S607, the determination unit 108 determines whether or not the strong classifier currently focused on is the last strong classifier (N-th strong classifier). If it is not the last strong classifier, the process proceeds to step S602, and the process proceeds to the next strong classifier discrimination process. In the case of the last discriminator, it is determined that the detection window currently focused on includes a person, and the flow proceeds to step S608, where the detection result position (x, y) of the detection window is stored in the determination result storage unit 109. Is stored.

[Subject likelihood calculation process (step S603)]
Next, the details of the subject likelihood calculation process in step S603 will be described. In conventional subject detection processing, subject likelihood calculation processing is performed at all positions in a detection window for raster scanning. However, such a process requires calculation of a very large number of subject likelihoods. On the other hand, there is a method of reducing the calculation cost by thinning out the scan pixels. However, when the scan pixels are simply thinned out, many detection omissions occur.

  In contrast, in the present embodiment, when calculating the subject likelihood, attention is paid to the shape of the local region corresponding to each weak discriminator, and the scan width is determined in accordance with this, so that the accuracy can be reduced at a high speed. Subject likelihood calculation is performed. 9 and 10 are diagrams showing the basic principle of subject likelihood calculation processing in the embodiment of the present invention.

  As shown in FIG. 9, when attention is paid to a local region 901 having a wide width and a local region 903 at a position where two pixels are skipped in the horizontal direction, a region overlapping with the number of skipped pixels Is big. For this reason, when the cumulative feature amount (for example, luminance histogram or HOG (Histgrams Of Oriented Gradients) feature) in each local region is represented as graphs 902 and 904, the difference between these features is small. The HOG feature is shown, for example, in “Navneet Dalal and Bill Triggs,“ Histograms of Oriented Gradients for Human Detection ”, IEEE Computer Vision and Pattern Recognition. Vol.1, pp.886-893, 2005”. The HOG feature uses a histogram based on the intensity and angle of the pixel gradient in the local region as a feature amount, and is a particularly effective feature amount for person detection.

  As shown in FIG. 10, when attention is paid to a local region 1001 having a wide vertical width and a local region 1003 at a position where the vertical region is skipped by two pixels, the number of skipped pixels is similarly overlapped. The area of the part to do is large. For this reason, when the cumulative feature amount in each local region is represented as graphs 1002 and 1004, the difference between these features is also small.

  Therefore, when performing the discrimination process using the cumulative feature quantity in the local area as the feature quantity, if the horizontal width of the local area to be discriminated is sufficiently wide, the horizontal step of the raster scan of the weak classifier Even if the width is thinned out, it can be said that the accuracy is not greatly affected. Similarly, if the vertical width of the local region to be discriminated is sufficiently wide, even if the step width in the vertical direction of the raster scan of the weak discriminator is thinned out, the accuracy is not greatly affected. .

  Therefore, in step S603, the following processing is executed. That is, in the present embodiment, when the width of the local area of the weak classifier is wide, raster scanning is performed by thinning out pixels in the horizontal direction (so that the horizontal scanning interval is widened). FIG. 11 is a flowchart showing the contents of the subject likelihood calculation process in step S603 in the first embodiment.

  First, in step S1101, the local feature amount calculation unit 105 performs the t th weak discriminator hnt (x, y) of the n th strong discriminator in the detection window (x, y) via the position control unit 106. The width value is acquired from the dictionary storage unit 107. When storing the weak classifier obtained as a result of learning in advance as a dictionary, the upper left X coordinate, the upper left Y coordinate, the width Wl, and the height Hl are stored as local area information as in the information 807 of FIG. Of these, the value of the width Wl is acquired.

  Next, in step S1102, the local feature amount calculation unit 105 determines whether or not the acquired value of the width Wl is larger than a predetermined threshold thw. If it is larger, it is determined that the region is a local region where pixel thinning is performed, and the flow proceeds to step S1103. Otherwise, the subject likelihood calculation is performed for each pixel, and the flow proceeds to step S1104.

In step S <b> 1103, the local feature amount calculation unit 105 determines whether or not the local region where pixel thinning is performed is in a position where pixel thinning is performed. In the present embodiment, since the pixel thinning number is set to 3 pixels, it is determined whether or not the following formula (1) is satisfied with respect to the horizontal coordinate x of the position (x, y) of the detection window.
x% 3! = 0 (1)
“%” Is calculated as a remainder.

  When the expression (1) is satisfied, the weak discriminator hnt (x, y) is in the thinning position, so that the calculation of the subject likelihood is omitted and the flow proceeds to step S1107.

  In step S1107, the local feature amount calculation unit 105 acquires the subject likelihood lnt by referring to the value held in step S1106.

  On the other hand, when Expression (1) is not satisfied, the weak discriminator hnt (x, y) is located at the position where the subject likelihood is calculated. Therefore, the flow proceeds to step S1104, and the local feature quantity calculation unit 105 performs the local feature calculation. The quantity Unt is calculated. Details of the local feature amount Unt calculation process in step S1104 will be described later.

Next, in step S1105, the local feature amount calculation unit 105 acquires the subject likelihood lnt. In obtaining the subject likelihood lnt, conversion from the local feature amount Unt is performed using the following equation (2).
lnt = fnt (Unt) (2)

  Here, the function fnt is a correspondence table that represents the relationship between the local feature amount and the subject likelihood in the t-th weak classifier hnt of the n-th strong classifier. The local feature quantity calculation unit 105 refers to the correspondence table and acquires the subject likelihood lnt from the local feature quantity Unt.

  Note that the subject likelihood lnt refers to this value instead of calculating the subject likelihood when the weak discriminator hnt is in the thinning position in the subsequent detection windows (step S1107). The data is stored in a memory provided in the amount calculation unit 105.

  Thereafter, in step S1108, the subject likelihood lnt is substituted for the subject likelihood Lnt (x, y) of the weak discriminator hnt, and the process returns to step S603 in FIG.

  In the above processing, for example, when the detection window scans in the horizontal direction of (0, y) → (1, y) → (2, y) → (3, y), weak detection is performed with the detection window (0, y). The object likelihood lnt of the container hnt is calculated and held. The subject likelihood calculation of the weak classifier hnt in the detection window (1, y) and the detection window (2, y) is omitted because it is in the thinning position, and the subject likelihoods Lnt (1, y) and Lnt (2, The subject likelihood lnt is assigned to both y). Then, when moving to the detection window (3, y), the subject likelihood is calculated again, and the value of the subject likelihood lnt is updated.

  In this example, it is determined whether or not to perform pixel thinning in the horizontal direction by referring to the horizontal width of the local region, but whether or not to perform pixel thinning in the vertical direction with reference to the vertical width of the local region is determined. It may be determined, or both may be combined. In other words, the vertical scanning interval may be determined to be wider as the vertical width of the local region is wider.

[Local Feature Quantity Calculation Processing (Step S1104)]
Next, details of the local feature amount calculation processing in step S1104 will be described. In the present embodiment, a great effect can be obtained by using a cumulative feature amount of pixel features as the local feature amount. Examples of the cumulative feature amount of the pixel feature include a feature amount obtained by accumulating pixel features such as the pixel value itself, the luminance value, or the edge strength in the local region, such as the HOG feature and the Haar feature. In this embodiment, HOG features that are particularly effective for human detection are used, but other feature amounts can also be used.

  In step S1104, the following processing is executed. FIG. 12 is a flowchart showing the contents of the local feature amount calculation process in step S1104.

  First, in step S1201, the local feature amount calculation unit 105 calculates gradient information. Here, the gradient information includes two pieces of information on the gradient strength and gradient direction of the pixel feature in the adjacent pixels. A luminance value is typical as a pixel feature, but any pixel feature may be used as long as it represents color information of other pixels. Here, the pixel feature is described as a luminance value. Let the luminance value at pixel (u, v) be I (u, v).

  FIG. 13 is a diagram illustrating the relationship between the gradient intensity and gradient direction of the luminance value related to the pixel of interest (u, v). The gradient intensity of luminance at the pixel (u, v) is represented by the following equation (3), and the gradient direction is represented by the following equation (4).

  The local feature amount calculation unit 105 calculates a value m (u, v) indicating the gradient strength 1301 and a value θ (u, v) indicating the gradient direction 1302 from these equations (3) and (4).

  Next, in step S1202, the local feature amount calculation unit 105 generates a gradient direction histogram. FIG. 14 is a diagram illustrating an example of a gradient direction histogram. Each pixel in the local region 1402 of the detection window 1401 in FIG. 14 has information on the gradient strength and gradient direction indicated by the arrow 1403 by the processing in step S1201. In step S1202, based on these pieces of information, a histogram weighted by the gradient strength for each gradient direction is generated. At this time, if the gradient direction from 0 degrees to 180 degrees is divided into nine stages, for example, a histogram 1404 is obtained. Hereinafter, the local region 1402 is referred to as a cell region. By the processing in step S1202, each cell region in the detection window 1401 has a nine-dimensional feature vector generated as a gradient direction histogram. A feature vector Fi in one cell is expressed by the following formula (5).

  Thereafter, in step S1203, local feature amount calculation section 105 normalizes the cell area in a block area composed of a plurality of cell areas. In the present embodiment, an area including a 2 × 2 cell area is set as a block area, and the 9-dimensional feature vector in the cell area is normalized with this block area. FIG. 15 is a diagram illustrating an example of a block area. Each cell in the block area 1501 has a 9-dimensional vector. If these are expressed as F1, F2, F3, and F4 for each cell area, a 36-dimensional vector exists in the block area. . Assuming that these are block feature vectors Vk, the block feature vectors Vk are expressed by the following equation (5).

  And the local feature-value calculation part 105 normalizes with respect to this block feature vector Vk by the following formula | equation (6) in order to reduce the illumination fluctuation | variation in pattern matching.

  The block feature vector Vk obtained as described above becomes the HOG feature in the local region. That is, in pattern discrimination of each weak discriminator, patterns are collated using 36-dimensional feature vectors.

  Here, the cell region is configured by 6 × 6 pixels, but the number of constituent pixels of the cell region is arbitrary, and the aspect ratio of the cell region is not limited to 1: 1. Similarly, the block area is composed of 2 × 2 cell areas, but the number of cell areas constituting the block area is arbitrary. In this way, by creating an arbitrary number of cell regions of various sizes and aspect ratios and generating block regions, it becomes possible to express the characteristics of multiple types of local regions having different shapes and sizes. .

  In such a first embodiment, when a person area is extracted as a subject in an image, raster scan pixel thinning is performed if the width of the local area corresponding to the discriminator is greater than or equal to a specific size. . For this reason, the discrimination process can be performed at high speed without degrading the discrimination performance.

(Second Embodiment)
Next, a second embodiment of the present invention will be described. The second embodiment mainly differs from the first embodiment in subject likelihood calculation processing in step S603, and the other configuration is the same as that of the first embodiment. FIG. 16 is a flowchart showing the contents of subject likelihood calculation processing in step S603 in the second embodiment of the present invention.

  First, in the same manner as in the first embodiment, in step S1101, the local feature amount calculation unit 105 passes the position control unit 106 via the position control unit 106 and the t-th of the n-th strong classifier in the detection window (x, y). The width value of the weak classifier hnt (x, y) is acquired from the dictionary storage unit 107.

  Next, in step S1601, the local feature amount calculation unit 105 determines the number of thinned pixels m based on the width Wl of the weak discriminator hnt. At this time, a correspondence table of the width of the weak classifier and the number of thinned pixels is prepared in advance, and the number m of thinned pixels is determined by referring to this table. FIG. 17 shows an example of the correspondence table when the detection window is 60 × 120.

Thereafter, in step S1602, the local feature amount calculation unit 105 determines the subject likelihood of the weak discriminator hnt in the detection window (x, y) from the horizontal coordinate x and the number of thinned pixels m of the currently focused detection window. It is determined whether to perform the calculation.
x% m! = 0 (7)

  When Expression (7) is satisfied, the weak discriminator hnt (x, y) is in the thinning position, and thus the calculation of subject likelihood is omitted and the flow proceeds to step S1107.

  In step S1107, the local feature amount calculation unit 105 acquires the subject likelihood lnt by referring to the value held in step S1106.

  On the other hand, if the expression (7) is not satisfied, the weak classifier hnt (x, y) is in a position where the subject likelihood is calculated, and therefore, the flow proceeds to step S1104 and the local feature quantity calculation unit 105 performs the local feature calculation. The quantity Unt is calculated.

  After step S1107 and after step S1104, the processes of steps S1105, S1106, and S1108 are performed in the same manner as in the first embodiment.

  In such a second embodiment, when extracting a person area as a subject in an image, the number of thinned pixels of a raster scan is determined according to the width value of the local area corresponding to the weak classifier, and the thinning is performed. Pixel thinning is performed by the number of pixels. For this reason, as in the first embodiment, the discrimination processing can be performed at high speed without degrading the discrimination performance.

  In the first and second embodiments, a person region in an image is extracted as a subject, but objects other than a person, such as a face and an animal, may be extracted.

  The processing of the above-described embodiment can also be realized by providing a system or apparatus with a recording medium that records a program code of software that embodies each function. The functions of the above-described embodiments can be realized by the computer (or CPU, MPU) of the system or apparatus reading and executing the program code stored in the recording medium. In this case, the program code itself read from the recording medium realizes the functions of the above-described embodiments, and the storage medium (computer-readable recording medium) storing the program code constitutes the present invention. become.

  In addition, a case where an OS (operating system) running on the computer performs part or all of the actual processing based on an instruction of the program code read by the computer is included.

  101: Image input unit, 102: Image storage unit, 103: Reduced image generation unit, 104: Reduced image setting unit, 105: Local feature amount calculation unit, 106: Position control unit, 107: Dictionary storage unit, 108: Determination unit 109: Determination result storage unit

Claims (9)

Dictionary storage means for storing dictionary information including information on the position and shape of a local region for determining a predetermined subject;
Position control means for controlling the position of the detection window in the image based on the dictionary information;
Determining means for determining a local region in the detection window based on the dictionary information;
Local feature amount calculating means for calculating a feature amount in the determined local region;
Determination means for determining whether or not the predetermined object is included in the detection window based on the feature amount and the dictionary information;
I have a,
The image processing apparatus according to claim 1, wherein the determining unit determines a position of the local region that is a target for calculating the feature amount such that a horizontal scanning interval becomes wider as a lateral width of the local region is larger . Said determining means according to claim 1, characterized in that to determine the position of the local region for which to calculate the characteristic amount such as longitudinal width is wider in the vertical direction of the scanning spacing becomes wider in the local region Image processing apparatus. Dictionary storage means for storing dictionary information including information on the position and shape of a local region for determining a predetermined subject;
Position control means for controlling the position of the detection window in the image based on the dictionary information;
Determining means for determining a local region in the detection window based on the dictionary information;
Local feature amount calculating means for calculating a feature amount in the determined local region;
Determination means for determining whether or not the predetermined object is included in the detection window based on the feature amount and the dictionary information;
Have
The image processing apparatus according to claim 1, wherein the determining unit determines a position of a local region that is a target for calculating the feature amount such that the vertical scanning interval becomes wider as the vertical width of the local region is larger. The feature amount, the image processing apparatus according to any one of claims 1 to 3, characterized in that a characteristic quantity of accumulating pixel feature of the local region. A position control step for controlling the position of the detection window in the image based on dictionary information stored in the dictionary storage means and including information on the position and shape of the local region for determining a predetermined subject;
Determining a local region in the detection window based on the dictionary information;
A local feature amount calculating step of calculating a feature amount in the determined local region;
A determination step of determining whether or not the predetermined subject is included in the detection window based on the feature amount and the dictionary information;
Have
In the determination step, the position of the local region for which the feature amount is calculated is determined such that the horizontal scanning interval becomes wider as the lateral width of the local region becomes wider . A position control step for controlling the position of the detection window in the image based on dictionary information stored in the dictionary storage means and including information on the position and shape of the local region for determining a predetermined subject;
Determining a local region in the detection window based on the dictionary information;
A local feature amount calculating step of calculating a feature amount in the determined local region;
A determination step of determining whether or not the predetermined subject is included in the detection window based on the feature amount and the dictionary information;
Have
In the determining step, the position of the local region for which the feature amount is calculated is determined such that the vertical scanning interval becomes wider as the vertical width of the local region becomes wider. On the computer,
A position control step for controlling the position of the detection window in the image based on dictionary information stored in the dictionary storage means and including information on the position and shape of the local region for determining a predetermined subject;
Determining a local region in the detection window based on the dictionary information;
A local feature amount calculating step of calculating a feature amount in the determined local region;
A determination step of determining whether or not the predetermined subject is included in the detection window based on the feature amount and the dictionary information;
Was executed,
Wherein in the determination step, the program characterized that you determine the position of the local region for which to calculate the characteristic amount such as lateral width is wide in the horizontal direction of the scanning spacing becomes wider in the local region. On the computer,
A position control step for controlling the position of the detection window in the image based on dictionary information stored in the dictionary storage means and including information on the position and shape of the local region for determining a predetermined subject;
Determining a local region in the detection window based on the dictionary information;
A local feature amount calculating step of calculating a feature amount in the determined local region;
A determination step of determining whether or not the predetermined subject is included in the detection window based on the feature amount and the dictionary information;
And execute
In the determination step, the position of the local region that is a target for calculating the feature amount is determined so that the vertical scanning interval becomes wider as the vertical width of the local region becomes wider. 9. A computer-readable recording medium on which the program according to claim 7 or 8 is recorded.

Images