JP2009140369A - Group learning device and group learning method, object detection device and object detection method, and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To perform excellent group learning of a group of weak discriminators comprising fairly free filters by using statistic learning, such as Adaboost. <P>SOLUTION: The weak discriminator is a four-point filter for calculating a feature amount of a gray image by inner product calculation between a luminance value vector having luminance values of reference pixels at four pixel positions as elements and a filter coefficient vector comprising arbitrary real number values. Learning of the weak discriminator is performed by determining a filter coefficient comprising a combination of pixel positions of reference pixels and an ideal real number value by using Logistic Regression so that a discrimination difference obtained by charging the weak discriminator with a plurality of learning samples is minimized. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a group learning device and a group learning method for collectively learning an object detection device for detecting whether or not a given image is a desired object such as a face, and a group learning target detection device And a group learning apparatus and a group learning method for collectively learning a classifier composed of a plurality of weak classifiers using statistical learning such as Adaboost, etc. The present invention relates to an object detection apparatus, an object detection method, and a computer program for detecting an object using a group of weak classifiers.

  More specifically, the present invention relates to a collective learning apparatus and collective learning method for collective learning of weak discriminator groups composed of filters having a high degree of freedom using statistical learning such as Adaboost, and weak discrimination obtained by collective learning. BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an object detection apparatus and an object detection method for detecting an object using a group of devices, and a computer program, and particularly ideal as a filter coefficient for multiplying an input image (a plurality of reference pixels) Collective learning apparatus and collective learning method for collective learning of a filter group having a high degree of freedom by obtaining real values, and an object detection apparatus and an object for detecting an object using a weak classifier group obtained by collective learning The present invention relates to a detection method and a computer program.

  Face recognition technology can be widely applied to man-machine interfaces, such as gender identification, as well as personal authentication systems that do not burden the user. Recently, face recognition has been applied to automate camera work such as automatic focusing (AF) and automatic exposure adjustment (AE) based on subject detection or subject recognition in a digital camera, automatic field angle setting, and automatic shooting. Yes.

  The face recognition system includes, for example, a face detection process that detects the position of a face image and extracts it as a detected face, a face organ detection process that detects the position of a main face organ from the detected face, Specific face) processing. In the face detection process, the size and position of the face are detected from the input image and extracted as a detected face. In the facial organ detection process, facial organs such as the center of the eyes, the eyes, the corners of the eyes, the nose and the eyebrows are found from the detected face. Then, after performing alignment and rotation correction based on the detected face organ position, the detected face is identified (person recognition, etc.) in the face identification process.

  FIG. 25 shows an example of a common image recognition procedure. As shown in the figure, filters such as a Haar like feature, a Gabor filter, and a steerable filter are widely used as preprocessing or feature extraction for image recognition. ing. This is because the feature extraction is often more advantageous than the direct discrimination of the image itself.

The filter referred to in this specification is expressed as the following formula (1). Filter Filter the luminance vector made up of the brightness value x _j of the reference pixels of a plurality of pixel locations, a dot product computation with coefficients vector composed of the filter coefficients w _j for multiplying each reference pixel. That is, the filter Filter is configured by a combination of a plurality (M in the following equation (1)) pixel positions used as reference pixels and filter coefficients to be applied (multiplied) to each reference pixel.

  The filter outputs whether or not a desired object (for example, a subject's face or smile) has been recognized from a given image, for example, with a positive or negative sign. One filter is only a “weak learner” (Weak Learner) that is slightly better than random (a weak discriminator uses some feature amount to determine whether it is an object or a non-object) ). In 1996, Freund et al. Proposed the theory that a strong classifier can be constructed by using a combination of a plurality of weak classifiers, that is, Adaboost. Each weak classifier is generated by placing a weight α on the weak classifier that the weak classifier generated before itself is not good at, and how reliable each weak classifier is, A majority decision is made based on this.

  However, the filter is usually designed in advance, and it is not obvious which filter is optimal to use from among many existing filters. Therefore, usually, a filter to be used is determined by a method such as a trial and error determination or a search for a brute force from a group of filters that can be discretely decomposed.

  The filter represented by the above formula (1) is composed of a combination of a plurality of pixel positions used as reference pixels and filter coefficients to be applied (multiplied) to each reference pixel. Which pixel position should be used for the reference pixel and what filter coefficient value is suitable for multiplying for each reference pixel is what to recognize from a given image (eg, whether to distinguish a face, smile It is necessary to determine the combination of the position of the reference pixel and the filter coefficient by trial and error.

  For example, a group learning device using a weak discriminator that uses a very simple feature quantity (difference feature between pixels), which is a difference in luminance value between two reference pixels, to determine whether or not an object is used as a filter is proposed. (For example, see Patent Document 1). The discriminator that uses the inter-pixel difference feature for the filter output can be expressed as the following equation (2).

A discriminator using the inter-pixel difference feature as a filter output is effective in that it can operate at a very high speed. However, as can be seen from the above equation (2), only two reference pixels are used, and the filter coefficients w ₁ and w ₂ applied to the two reference pixels are fixed to +1 and −1. Due to restrictions, the degree of freedom as a filter is very low, and the recognition performance is sacrificed.

In the above equation (2), if the filter coefficient w _j can be selected not only from ± 1 but also from an arbitrary real number, the degree of freedom as a filter increases, so that the recognition performance may be improved. However, it is not obvious how to determine the coefficient consisting of the real number. For this reason, a fixed coefficient is used in the pixel difference feature.

  In the Sparse feature proposed by Huang et al., An effective filter is constructed by combining various Haar-like features (adding (+) or subtracting (-) filters for the average pixel value of a rectangular area in an image). (See, for example, Non-Patent Document 1). However, this method corresponds to searching for a space composed of ternary coefficients {-1, 0, +1}, in other words, using a fixed filter coefficient instead of an arbitrary real number. The degree of freedom of the filter that can be originally set is not used up.

  The definition of the sparse feature is as shown in the following formula (3). Expressing power can be enhanced by making α in the formula not a binary value of {−1, +1} but a real number. However, since it is not usually known which real value the filter coefficient is set to be advantageous for the discriminator at the subsequent stage, the search is limited to discrete values. In other words, the present inventors consider that the recognition performance can be improved if an ideal real value used as a filter coefficient can be obtained.

  In the above equation (3), a pixel that is not referred to can be considered as a filter composed of ternary coefficients because α can be considered to be 0. In FIG. 26, the black square corresponds to α = −1, the white square corresponds to α = + 1, and the gray corresponds to α = 0.

  If you can automatically create a filter that will be advantageous for the target recognition task, you will not only be freed from the effort of trial and error, but also contribute to improving recognition performance, reducing processing volume and speeding up. Can be expected and is very effective.

JP 2005-157679 A Huang, C.I. , Ai, H .; Li, Y .; , And Lao, S .; 2006. "Learning Sparse Features in Granular Space for Multi-View Face Detection" (In Proceedings of the 7th international Conference on Automatic Face and Gesture Recognition (April 10-12,2006) .FGR.IEEE Computer Society, Washington, DC, 401-407 )

  An object of the present invention is to provide an excellent group learning device and group learning that can suitably perform group learning for an object detection device for detecting whether or not a given image is a desired object such as a face. An object is to provide a method, a group-learned object detection device and an object detection method, and a computer program.

  A further object of the present invention is to obtain an excellent group learning apparatus and group learning method and group learning, which can suitably perform group learning of a classifier consisting of a plurality of weak classifiers using statistical learning such as Adaboost. Another object of the present invention is to provide an object detection apparatus, an object detection method, and a computer program that detect an object using a group of weak classifiers.

  A further object of the present invention is to provide an excellent group learning apparatus, group learning method, and group learning that can suitably perform group learning of weak classifier groups composed of filters having a high degree of freedom using statistical learning such as Adaboost. Another object of the present invention is to provide an object detection device, an object detection method, and a computer program for detecting an object using the weak classifier group obtained in this way.

  A further object of the present invention is to obtain an ideal real value to be used as a filter coefficient for multiplying an input image (a plurality of reference pixels) and to perform group learning suitably for a filter group having a high degree of freedom. To provide a group learning apparatus and a group learning method, an object detection apparatus and an object detection method for detecting an object using a weak classifier group obtained by group learning, and a computer program.

The present invention has been made in consideration of the above problems, and the first aspect of the present invention is to use a learning sample consisting of a plurality of grayscale images that are known to be objects or non-objects. A collective learning apparatus that collectively learns a plurality of weak discriminators used for detecting an object for detecting whether or not a given image is an object,
Each weak discriminator is a filter that calculates a feature amount by calculating an inner product of a luminance value vector whose element is the luminance value of a reference pixel at the pixel position of the L point and a filter coefficient vector consisting of an arbitrary real value (however, L is an integer of 2 or more),
For each weak discriminator, a learning means for learning a combination of reference pixel positions and an ideal real value used for a filter coefficient by boosting is provided.
This is a group learning device characterized by this.

The second aspect of the present invention is an object detection device for detecting whether or not a given grayscale image is an object,
A plurality of filters each including a filter for calculating a feature value of the given grayscale image by calculating an inner product of a luminance value vector having the luminance value of the reference pixel at the pixel position of the point L as an element and a filter coefficient vector composed of an arbitrary real value. A weak classifier,
A discriminator for discriminating whether or not the given grayscale image is an object based on the feature amount calculated by at least one of the plurality of weak discriminators;
It is an object detection apparatus characterized by comprising.

  Filters are widely used as preprocessing or feature extraction for image recognition. The filter is an inner product calculation of a luminance vector composed of the luminance values of the reference pixels at a plurality of pixel positions in the input image and a coefficient vector composed of a filter coefficient multiplied by the luminance value of each reference pixel. One filter is only a weak classifier that is slightly better than random, but a strong classifier can be constructed by using a combination of a plurality of weak classifiers.

  Conventional filters use only fixed filter coefficients, do not fully use the degree of freedom of the filter, and sacrifice recognition performance. If an ideal real value used as a filter coefficient can be obtained, recognition performance can be improved.

  On the other hand, according to the present invention, it is possible to obtain an ideal real value used as a filter coefficient by reducing learning of a filter to a linear two-class discrimination problem, and to appropriately learn a filter group having a high degree of freedom. Can do.

  The learning means inputs a plurality of learning samples consisting of gray images in which two classes of objects or non-objects are classified, that is, labeled, into each weak classifier, and each of the objects and non-objects by boosting The feature amount of is learned in advance.

  Specifically, the learning means includes a multi-point filter including a combination of the pixel positions of the L points of the reference pixels and a filter coefficient so as to minimize a discrimination error obtained by inputting a plurality of learning samples into the weak discriminator. The weak discriminator is learned by determining the parameters. Such a minimization problem can be solved, for example, using algorithms such as Logistic Regression or Support Vector Machine. The Logistic Regression or Support Vector Machine approach itself is well known in the art.

More specifically, the learning means initializes the data weights D _{1, i} of each learning sample for the weak classifier to be generated first,

Weak discriminator learning that learns a weak discriminator using Logistic Regression or a support vector machine, and each misclassified learning when a plurality of learning samples are input to the learned t-th weak discriminator A weighting error calculation for adding a sample data weight D _{t, i} to calculate a weighting error, and a weak classifier used for weighted majority determination of feature quantities calculated by each weak classifier at the time of detection of an object. Weight calculation of weighted majority voting to calculate weight α _t , data weight D _{t + 1, i} of each learning sample, weight voting weight α _t , and learning result for each learning sample by learned weak classifier It is possible to learn the necessary number of weak discriminators by repeating the data weight update calculated in the above manner as many times as necessary.

  Here, in the weak discriminator learning process, the combination of reference pixel positions of the multipoint filter t is determined at random, and the filter coefficient of the multipoint filter t that minimizes the cost function including the weighting error is Logistic Regression or supported. The determination using the vector machine and the process for obtaining the weighting error of the multipoint filter t having the determined filter coefficient are repeated a plurality of times. Then, a multi-point filter t composed of a combination of reference pixel positions and a filter coefficient that minimizes a weighting error in this iterative process may be adopted as the t-th weak discriminator.

  Therefore, according to the present invention, it is possible to automatically create a filter that is advantageous for the target recognition task, so that not only the labor of trial and error is released, but the obtained weak classifier group is It can be expected to contribute to improvement of recognition performance, reduction of processing amount, and speedup in the applied object detection device, which is very effective.

  Here, an input image such as a learning sample to be handled at the time of learning of the weak classifier and an image given at the time of detection of the target is cut out to a predetermined window size such as 20 × 20 pixels. Each weak discriminator arranges L reference pixels in a straight line on the window size, and holds each pixel position as two pieces of information of {start position q, step width s}. Good. That is, by holding the pixel position information in a compact representation as much as possible, the number of memory accesses during execution such as learning or detection can be reduced, and the processing speed can be increased.

  In addition, a 2 × 2 blurring process is performed on an input image of a 1 × 1 layer, such as a learning sample or a grayscale image to be examined, by averaging luminance values for each adjacent 2 × 2 pixel block to 1 pixel. A 2 × 2 layer image is generated, and the original learning sample is subjected to 4 × 4 blurring processing by averaging the luminance value for each adjacent 2 × 2 pixel block to 1 pixel, and the 4 × 4 layer Multi-layer image generation means for generating an image may be further provided. Each weak classifier uses a learning sample of each layer that has been made multi-layered, and a multi-point filter for any one of the layers learns in advance. In such a case, at the time of detecting an object, it is possible to calculate a feature amount for a corresponding layer obtained by multi-layering the given grayscale image by the multi-layer image generating means.

  In each of the 2 × 2 and 4 × 4 layers, the multipoint filter is obtained by arranging enlarged L-point reference pixels having sizes of 2 × 2 and 4 × 4 with respect to the original 1 × 1 pixel. . Referencing each pixel of the blurred image is equivalent to referencing a wide range at a time in the original image of the 1 × 1 layer.

  Depending on the position of the face part, the use of a 1 × 1 layer multipoint filter that refers only to a fine and narrow range of granularity provides better recognition performance, or conversely, a 2 × 2 layer that refers to a wide range with a coarse granularity. Or the case where the multipoint filter of 4x4 layer uses the recognition performance becomes higher. In addition, when a large number of pixels are referred to, the amount of calculation increases. However, the number of reference pixels can be reduced as much as possible by making the image multi-layered.

  Further, at the time of detection of an object, it is determined that the given grayscale image is not an object according to a comparison result between an in-computation value for weighted majority determination of the feature amount calculated by each weak discriminator and an abort threshold. Thus, it is possible to introduce an abort process for aborting the calculation.

  This censoring threshold can be calculated based on the minimum value that can be taken by the weighted majority value of the discrimination result of the learning sample indicating the object. When the weak discriminator learning, the weighting error calculation, the weight calculation, and the data weight update process are repeated as many times as necessary, the abort threshold value for each weak discriminator is obtained by performing the abort threshold value calculation process together. Can learn.

  Each time a weak classifier outputs a feature value and updates the value of the weighted majority vote, it is compared with the abort threshold value, and the window image is not an object when the weighted majority vote value falls below the abort threshold value. By aborting the calculation, it is possible to speed up the discrimination process without using unnecessary calculations.

The third aspect of the present invention uses a learning sample consisting of a plurality of grayscale images whose correct answers (whether they are objects or non-objects), and the given image is the object. A computer program written in a computer-readable format so as to execute on a computer a process for collective learning of a plurality of weak classifiers used for detecting an object for detecting whether or not there is an object,
Each weak discriminator is a multi-point filter that calculates a feature value by calculating an inner product of a luminance value vector whose element is the luminance value of a reference pixel at the pixel position of the L point and a filter coefficient vector composed of an arbitrary real value ( However, L is an integer of 2 or more, and each learning sample i has a data weight D _{t, i} that represents a weight reflecting the difficulty of discrimination for each weak discriminator t (where i identifies a learning sample) Serial number, t is a serial number that identifies the weak classifier),
The computer program causes the computer to
Initialization means for initializing data weights D _{1, i} of each learning sample for the weak classifier to be generated first;
Weak classifier learning means for learning the t-th weak classifier by using Logistic Regression or a support vector machine;
When a plurality of learning samples are input to the t-th weak classifier learned by the weak classifier learning means, a data error D _{t, i} of each misclassified learning sample is added to calculate a weighting error. A weighting error calculating means;
A weighted majority voting weight calculating means for calculating a weight α _t corresponding to the reliability of the t-th weak classifier, which is used when the weighted majority of the feature quantities calculated by each weak classifier at the time of detection of an object;
The data weight D _{t + 1, i} of each learning sample is calculated based on the weighting majority weight α _t , the label for correctly attaching each learning sample, and the discrimination result for each learning sample by the t-th weak discriminator. Data weight updating means to perform,
Function as
The weak classifier learning means, the weighting error calculation means, the weight calculation means, and the data weight update means repeat the process as many times as necessary to learn the required number of weak classifiers.
This is a computer program characterized by the above.

The fourth aspect of the present invention is a computer program written in a computer-readable format so that processing for detecting whether a given grayscale image is an object is executed on the computer. The computer
It consists of a multi-point filter that calculates the feature value of the given grayscale image by calculating the inner product of a luminance value vector whose element is the luminance value of the reference pixel at the L pixel position and a filter coefficient vector consisting of an arbitrary real value. A plurality of weak discrimination means (where L is an integer of 2 or more),
Discriminating means for discriminating whether or not the given grayscale image is an object based on the feature amount calculated by at least one of the plurality of weak discriminating means;
It is a computer program for making it function as.

  The computer program according to each of the third and fourth aspects of the present invention defines a computer program written in a computer-readable format so as to realize predetermined processing on the computer. In other words, by installing the computer program according to each of the third and fourth aspects of the present invention in a computer, a cooperative action is exhibited on the computer, and the group learning according to the first aspect of the present invention. Effects similar to those of the apparatus and the object detection apparatus according to the fourth aspect can be obtained.

  Advantageous Effects of Invention According to the present invention, an excellent group learning device and group learning that can suitably perform group learning for an object detection device for detecting whether or not a given image is a desired object such as a face. A method, a group-learned object detection device and an object detection method, and a computer program can be provided.

  In addition, according to the present invention, an excellent group learning apparatus and group learning method that can suitably perform group learning of a classifier composed of a plurality of weak classifiers using statistical learning such as Adaboost, obtained by group learning. It is possible to provide an object detection device, an object detection method, and a computer program that detect an object using the weak classifier group.

  In addition, according to the present invention, an excellent group learning apparatus and group learning method, a group learning method, and a group learning method that can suitably perform group learning of a weak classifier group including a filter having a high degree of freedom using statistical learning such as Adaboost. It is possible to provide an object detection apparatus, an object detection method, and a computer program for detecting an object using a weak classifier group obtained by learning.

  In addition, according to the present invention, it is possible to appropriately collect and learn a filter group having a high degree of freedom by obtaining an ideal real value to be used as a filter coefficient for multiplying an input image (a plurality of reference pixels). A group learning apparatus and group learning method, an object detection apparatus and an object detection method for detecting an object using a weak classifier group obtained by group learning, and a computer program can be provided.

  In addition, by configuring the object detection device using the weak classifier group obtained by the group learning device according to the present invention, it is possible to perform image recognition efficiently and at high speed with small data, and pixel difference Higher recognition performance can be expected than when weak classifier groups for obtaining features are used.

  Other objects, features, and advantages of the present invention will become apparent from more detailed description based on embodiments of the present invention described later and the accompanying drawings.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  In an object detection apparatus that detects an object from an image, a filter is widely used as preprocessing for discrimination or feature extraction. Such a filter can be obtained through ensemble learning (group learning).

  A learning machine obtained by group learning includes a number of weak hypotheses and a combiner that combines them. Here, boosting is an example of a combiner that integrates weak hypothesis outputs with fixed weights regardless of input. Boosting uses the previously generated weak hypothesis learning result to increase the weight of the learning sample that is not prone to mistakes, and then processes the distribution that the learning sample follows and learns a new weak hypothesis based on this distribution. To do. As a result, the weight of the learning sample that has many incorrect answers and is difficult to discriminate as a target is relatively increased, and a weak discriminator that makes the learning sample that has a large weight, that is, difficult to discriminate, correct is selected sequentially. The generation of weak hypotheses in learning is performed sequentially, and the weak hypotheses generated later depend on the weak hypotheses generated before that.

  When detecting an object, discrimination results of a number of weak discriminators sequentially generated by learning are used. For example, in the AdaBoost algorithm, the discrimination results of all weak discriminators generated by learning (1 for an object, -1 for a non-object) are supplied to the coupler, and the coupler for each weak discriminator. A weighted majority result obtained by weighting and adding the reliability calculated at the time of learning is output, and the output value is evaluated to select whether or not the input image is an object.

  The weak classifier is a filter that determines whether the object is an object or a non-object by using some feature amount. In the following, it is assumed that the filter performs an inner product calculation of a luminance vector composed of the luminance values of reference pixels at a plurality of pixel positions in the input image and a coefficient vector composed of a filter coefficient to be multiplied to each reference pixel (the above formula (See (1)). The output of the weak classifier may be output deterministically whether or not the object is an object, and may be output probabilistically by the probability density or the like.

  Patent Document 1 (described above) uses a group learning device that uses a very simple weak classifier that determines whether or not an object is an object from a difference in luminance value between two pixels, that is, an inter-pixel difference feature. There has been proposed a method for speeding up the object detection process while sacrificing. On the other hand, in this embodiment, the learning performance of the filter is reduced to a linear two-class discrimination problem, an ideal real value used as a filter coefficient is obtained, and the recognition performance is improved by using a filter group having a high degree of freedom. We are trying to improve. The method of configuring the weak classifier, that is, the filter will be described in detail later.

Device configuration:
FIG. 1 schematically shows a functional configuration of the object detection apparatus. The illustrated object detection apparatus 1 includes an image input unit 2, a scaling unit 3, a scanning unit 4, and a discriminator 5.

  The image input unit 2 inputs a grayscale image (luminance image) captured by a digital camera, for example. The scaling unit 3 outputs a scaled image obtained by enlarging or reducing the input image to all designated scales. For each scaled image, the scanning unit 3 sequentially scans a window having a size of an object to be detected, for example, sequentially from the top line to the bottom, and cuts out the window image. Then, the discriminator 5 discriminates whether each window image sequentially scanned by the scanning unit 4 is an object such as a face or a non-object, and when the object is detected, the position and size indicating the area are determined. Output as detection result.

  Here, the group learning machine 6 performs group learning of a plurality of weak classifiers constituting the classifier 5 by group learning. The discriminator 5 refers to the learning result of the group learning machine 6 and discriminates whether the current window image is an object such as a face image or a non-object. The group learning machine 6 may be a component in the object detection device 1 or an external independent device.

  Moreover, the target object detection apparatus 1 outputs a plurality of region information when a plurality of target objects are detected from the input image. Furthermore, when there is an overlapping area among a plurality of area information, it is possible to perform a process of selecting an area having the highest evaluation as an object.

  The image (grayscale image) input to the image input unit 2 is first supplied to the scaling unit 3. The scaling unit 3 performs image reduction using bilinear interpolation. Instead of first generating a plurality of reduced images, the necessary image is output to the scanning unit 4, and after the processing of the image is completed, the next smaller reduced image is generated. repeat. FIG. 2 shows how the scaling unit 3 generates a reduced image. As shown in the figure, the input image 10A is output to the scanning unit 4 as it is, and after the processing of the scanning unit 4 and the discriminator 5 is completed, the input image 10B in which the size of the input image 10A is reduced is generated. . Subsequently, after the processing of the scanning unit 4 and the discriminator 5 in the input image 10B is completed, the input image 10C obtained by further reducing the size of the input image 10B is output to the scanning unit 4 so that the reduced image 10D, 10E etc. are generated. Then, the process ends when the image size of the reduced image becomes smaller than the window size scanned by the scanning unit 4. The image input unit 2 outputs the next input image to the scaling unit 3 after such processing is completed.

  FIG. 3 shows how the scanning unit 4 scans a window S having a predetermined window size on the input image. The window size is a size that is accepted by the discriminator 5 at the subsequent stage (that is, suitable for discrimination of an object), and is, for example, 20 × 20 pixels. The scanning unit 4 sequentially applies the window S to the entire image (screen) with respect to the input image 10 </ b> A, and outputs an image at each scan position (hereinafter referred to as a “cut image”) to the discriminator 5. . Although the window size S is constant, as shown in FIG. 2, the scaling unit 3 sequentially reduces the input image and scales it to various image sizes, so that it is possible to detect a target object of any size. It becomes.

The discriminator 5 determines whether or not the cut image provided from the scanning unit 4 is a target object such as a face. FIG. 4 shows the configuration of the discriminator 5. The discriminator 5 includes a plurality of weak discriminators 21 _t (21 _{1 to} 21 _K ) and an adder 22 that multiplies these outputs by weights α _t (α _{1 to} α _K ), respectively, and obtains a weighted majority vote. (T is a serial number for identifying weak classifiers, and K is the total number of weak classifiers). The discriminator 6 sequentially outputs estimated values as to whether or not each weak discriminator 21 _t (21 _{1 to} 21 _K ) is an object for the input window image, and the adder 22 calculates a weighted majority vote. And output. The discriminator 5 can determine whether or not the cut image is an object according to the weighted majority value.

Each weak discriminator 21 _t (21 _{1 to} 21 _K ) constituting the discriminator 5 multiplies each reference pixel by a luminance value vector composed of luminance values of reference pixels at a plurality of pixel positions in the input cut image. This is a filter for calculating a feature value for discrimination by calculating an inner product with a coefficient vector composed of the filter coefficients to be combined, and can be obtained by ensemble learning (Ensemble Learning). In this embodiment, the learning of the filter is reduced to a linear two-class discrimination problem, an ideal real value used as a filter coefficient is obtained, and the recognition performance is improved by using a filter group having a high degree of freedom. Yes.

The collective learning machine 6 learns in advance by collective learning a weak discriminator 21 _t (21 _{1 to} 21 _K ) and weights to be multiplied by their outputs (estimated values). As group learning, a technique that can obtain the results of a plurality of discriminators by majority decision can be applied. For example, group learning using boosting such as AdaBoost that performs weighted majority by weighting data can be applied.

At the time of learning, a plurality of learning samples consisting of gray images obtained by classifying or labeling two classes of objects or non-objects are input to the weak classifier 21 _t (21 _{1 to} 21 _K ), The feature amount of each of the object and the non-object is learned in advance. In the determination, the feature amount calculated for the window image sequentially supplied from the scanning unit 4 is compared with the feature amount of each of the target object and the non-target object learned in advance, and the window image is the target. An estimated value for estimating whether or not it is an object is output deterministically or probabilistically.

The adder 22 adds a weight α _t (α ₁ ) that is a reliability for each weak classifier 21 _t (21 _{1 to} 21 _K ) to the feature amount calculated from each weak classifier 21 _t (21 _{1 to} 21 _K ). to? _K) by multiplying the outputs the addition value (value of weighted majority decision) this. In AdaBoost, a plurality of weak classifiers 21 _t (21 _{1 to} 21 _K ) sequentially calculate estimated values, and the weighted majority values are sequentially updated accordingly. The plurality of weak classifiers 21 _n (21 _{1 to} 21 _N ) are sequentially generated by the group learning machine 6 by group learning using learning samples, and for example, feature quantities are calculated in the generation order. Also, the weighted majority vote weight α _t (α _{1 to} α _K ) (reliability) is learned in a learning process for generating each weak classifier 21 _t (21 _{1 to} 21 _K ).

When the weak discriminator 21 _t (21 _{1 to} 21 _K ) should perform binary output as in AdaBoost, the weak discriminator 21 _t (21 _{1 to} 21 _K ) It is determined whether it exists. Of course, you may make it discriminate | determine using a some threshold value. Further, the weak classifier 21 _t (21 _{1 to} 21 _K ) may probabilistically output a continuous value indicating the degree of being a target object from the feature amount calculated as Real-AdaBoost. The feature quantity (threshold value) for discrimination required by these weak discriminators 21 _t (21 _{1 to} 21 _K ) is also learned according to an algorithm such as AdaBoost at the time of learning.

When a weighted majority decision is made by the adder 22, the calculation result of all weak discriminators 21 _t (21 _{1 to} 21 _K ) is not waited, and the calculation is aborted by determining that the object is not a target object depending on the value during the calculation. It may be. With this abort process, the amount of calculation in the detection process can be significantly reduced. As a result, it is possible to move to the next window image discrimination process during the calculation without waiting for the calculation results of all the weak discriminators. The censoring threshold value can be learned at the time of learning.

Thus, the discriminator 5 calculates a weighted majority value as the evaluation value s for determining whether or not the window image is an object, and determines whether or not the window image is an object based on the evaluation value. judge. Further, each time the plurality of weak discriminators 21 _t (21 _{1 to} 21 _K ) sequentially calculate and output feature amounts, the discriminator 5 each weak discriminator 21 _t obtained by learning with respect to the feature amount. The weighted majority value obtained by multiplying and adding the weights for (21 _{1 to} 21 _K ) is updated. Then, every time the weighted majority value (evaluation value) is updated, it is possible to control whether or not the calculation of the estimated value is aborted using the abort threshold.

Weak classifier group learning:
As described above, the group learning machine 6 generates the discriminator 5 by performing group learning using a learning sample according to a predetermined group algorithm. Below, the structure of the weak discriminator 21 used with the discriminator 5 is demonstrated first, and the group learning method in the group learning machine 6 is demonstrated continuously.

  The group learning machine 6 obtains a learning sample consisting of a grayscale image in which whether it is an object or a non-object (for example, a face image or a non-face image) is correctly labeled (labeled) in advance. The weak classifier is generated by selecting (learning) one hypothesis from a large number of learning models (hypothesis combinations) according to a predetermined learning algorithm, and the combination of the weak classifiers is determined.

  Usually, learning is performed using thousands or tens of thousands of learning samples. FIG. 5 illustrates sample images of face images and non-face images. In this specification, a 400-dimensional luminance value vector is x, and a class label that classifies two classes of face and non-face is y.

The group learning machine 6 performs group learning using a boosting algorithm, but combines a plurality of weak classifiers and learns so that a strong determination result is obtained as a result. Each weak discriminator has low discrimination ability (a little better than random), but for example, hundreds to thousands of weak discriminators are selected, and as a result, they have high discrimination ability. A classifier can be constructed. The group learning machine 6 selects the weak discriminators, that is, selects the weak discriminators 21 _t (21 _{1 to} 21 _K ), and the weights α _t (α _{1 to} α when the output values are subjected to a weighted majority decision. _K ) to learn.

  In the present embodiment, learning of a filter is reduced to a linear two-class discrimination problem, whereby an ideal real value is obtained as a filter coefficient, and a weak discriminator composed of a filter having a high degree of freedom is learned. And the recognition performance is improved by using such a weak classifier group.

The weak classifier disclosed in Patent Document 1 (described above) is a so-called two-point filter that calculates a difference in luminance value between two pixels, that is, an inter-pixel difference feature. FIG. 6 shows how the difference feature between pixels is obtained. The two-point filter is composed of luminance values {I ₁ , I ₂ } at _two pixel positions 31 ₁ and 31 ₂ used as reference pixels and fixed filter coefficients {+1, −1}. 6 learns how to select and combine a number of two-point filters.

  In contrast, in the present embodiment, a filter using three or more pixels as reference pixels is used for the weak classifier. If the number of reference pixels increases, the recognition performance improves, but the calculation amount increases. In the following description, a case where a four-point filter having four reference pixels is used as a weak classifier is taken as an example. In order to configure a four-point filter, it is necessary to determine, as parameters, a combination of positions of four reference pixels and a real-value filter coefficient that multiplies the luminance value of each reference pixel. The collective learning machine 6 learns how to select and combine a large number of four-point filters, but obtains a filter coefficient composed of ideal real values by reducing it to a linear two-class discrimination problem, and a filter group having a high degree of freedom. To learn.

FIG. 7 shows a configuration example of a four-point filter. In the figure, four reference pixels are arranged in a straight line on an input image composed of 20 × 20 pixels (in the example shown, a face image as a learning sample). The illustrated filter has only two pieces of information of {start position q, step width s} to represent the positions of the four reference pixels, and has an advantage that the data to be held becomes compact. The position p _j of four reference pixels (where j is a serial number of the reference pixel and an integer of 1 to 4) is expressed by the following expression (4).

  As can be seen from FIG. 7, in the case of an input image consisting of 20 pixels per line, if an arbitrary start position q is set and the step width s is set to 1, the 4-point filter consists of 4 pixels in a horizontal row, and the step width s When 20 is set, the 4-point filter is composed of 4 pixels in a vertical column. When the step width s is 21, the 4-point filter is composed of 4 pixels slanting to the right. When the step width s is 19, the 4-point filter is increasing to the right. It consists of four diagonal pixels.

  Further, as shown in FIGS. 8A and 8B, a four-point filter that can refer to a wide range can be configured by performing hierarchical blurring processing on the input image. In FIG. 8A, the blurring process is performed by averaging the luminance values for each adjacent 2 × 2 pixel block to one pixel. In FIG. 8B, the blurring process is performed by averaging the luminance values for each adjacent 4 × 4 pixel block to one pixel. In these cases, the four-point filter is obtained by linearly arranging four enlarged reference pixels having a size of 2 × 2 and 4 × 4 with respect to the original 1 × 1 pixel. Referencing each pixel of the blurred image is equivalent to referencing a wide range at a time in the original image of the 1 × 1 layer.

  Depending on the position of the face part, the use of a 1 × 1 layer four-point filter that refers only to a fine and narrow range of the granularity provides better recognition performance, or conversely, a 2 × 2 layer that refers to a wide range with a coarse granularity. Alternatively, it is conceivable that the recognition performance is higher when a 4 × 4 layer four-point filter is used.

  In the present embodiment, as shown in FIG. 9, the original input image of 1 × 1 layer is a combination of 2 × 2 layer and 4 × 4 layer blurred image processed as a single input image. This means that an appropriate layer can be selected by designating the start position of the four-point filter.

  A processing module for processing the input image into a multi-layer can be disposed, for example, in the subsequent stage of the scanning unit 4 or incorporated in the scanning unit 4.

  The four-point filter not only simply increases the number of reference pixels, but also multiplies (internal product calculation) a luminance pixel vector x composed of luminance values of the respective reference pixels with respect to the pixel difference feature described in Patent Document 1. A characteristic is that the filter coefficient to be obtained is not a fixed value such as {+1, -1} but an ideal real value is obtained by collective learning. Thereby, a filter group with a high degree of freedom can be configured.

Filter f is a luminance image vector x L-dimensional, when there is a position p _j of the reference pixels of the filter is expressed as the following equation (5). However, the position p _{j of the} reference pixel is represented by {start position q _j , step width s _j } (see the above-mentioned and expression (4)). When the filter f is a four-point filter, the dimension number L = 4.

In the above equation (5), w _j is a filter coefficient for multiplying each element of the luminance pixel vector x, and b is a bias. The filter coefficient is not a fixed value such as {+1, −1} but an ideal real value. An ideal filter coefficient consisting of real values can be obtained by reducing to a linear two-class discrimination problem.

  The weak discriminator h using the filter f shown in the above formula (5) is the following formula (6), and the discrimination result is {+1, -1}, that is, the class label for classifying the two classes is y. Output (the inner product calculation result by the filter f is a positive value when an object is detected in the input image (for example, for a face image), and is not detected (for example, for a non-face image) Is negative).

Introduced as a virtual fifth reference pixels p ₅ as in the following equation (7), when the bias b included in the right side of the above equation (5) and the luminance value x _p5 of the virtual reference pixels p ₅ , The bias b can be expressed as one of the filter coefficients w ₅ , and the above equation (5) can be simplified to a vector inner product calculation form as shown in equation (8).

Assume that N sets of luminance value vectors (feature vectors) x and a class label y for classifying the discrimination results of the two classes are given by the following equation (9). The luminance value vectors x ⁽¹⁾ ,..., X ^(N) are image samples extracted in total from N learning samples (see FIG. 5 ⁾ of a large number of face images and non-face images. Is a vector whose elements are luminance values of four reference pixels. The class label y is +1 for the luminance value vector extracted from the face image sample, and the class label y is −1 for the luminance value vector extracted from the non-face image sample.

The discrimination error e _lr (= negative log likelihood: negative log-likelihood) at this time is expressed by the following equation (10). The collective learning machine 6 performs learning so as to select and combine a plurality of weak classifiers h that minimize the discrimination error _elr .

The parameter {w _j , b} of the weak discriminator h that minimizes the discrimination error e _lr is obtained, that is, the learning of the filter is realized by reducing it to a linear two-class discrimination problem. For this, it is possible to use Logistic Regression.

Further, determination error e _L2SVM and parameters of weak classifiers h which minimizes the _e l1SVM {w _j, b} are represented by the following formula (11) and (12) to seek a linear support vector machine ( Support Vector Machine) (L2-SVM and L1-SVM, respectively) can be used.

  The Logistic Regression or Support Vector Machine approach itself is well known in the art. For details of Logistic Regression, see, for example, Lin, C. et al. , Weng, R .; C. , And Keerthi, S .; S. 2007. Trust region Newton methods for large-scale logistic regression. (In Processeds of the 24th international, E. M.7.E., 2). ACM, New York, NY, 561-568). For details of the linear SVM, see, for example, Joachims, T .; 2006. Training linear SVMs in linear time. (See In Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Philadelphia, PA, USA, August 20-23, 2006). .

Note that the difference between Logistic Regression and SVM is only the above-mentioned cost function, and it is known that the performance is very close (for example, in the former Lin literature, Logistic Regression is the same as SVM) It is described that). Both techniques are well studied in the industry, and there are very fast learning algorithms in recent years. In addition, by using any of the methods, an ideal real value can be learned as the filter coefficient w _j of the weak classifier, and the classifier 5 including the weak classifier group having a high degree of freedom is configured to recognize the performance. It should be fully understood that this can be improved.

  The learning method of the weak classifier will be described in more detail. FIG. 10 shows boosting (Adaboost), which is a framework for learning. However, N (for example, several tens of thousands) learning samples X and learning sample labels (correct answers) are known. The learning sample consists of a large number of face images and non-face image sample images (see FIG. 5).

In the AdaBoost algorithm, a weak weight discrimination is performed such that a data weight D representing a weight reflecting the difficulty of discrimination is attached to each learning sample, and the feature quantity calculated by the four-point filter as the weak discriminator minimizes the weight error. The instrument parameters are learned. The parameters referred to here are a combination of pixel positions of reference pixels {p ₁ , p ₂ , p ₃ , p ₄ } and filter coefficients {w ₁ , w ₂ , w _{3 which} are necessary for constructing a filter. , W ₄ } and a bias w _{5 (} expressed as a filter coefficient of a virtual fifth reference pixel).

For the t-th weak classifier 21 _t , the data weight of the i-th learning sample is D _{t, i} . However, since the difficulty of discrimination of each learning sample is unknown at the initial stage, the following equation (14) is used as the data weight D _{1, i} of each learning sample for the weak discriminator 21 ₁ generated first by learning. ), A uniform value obtained by dividing 1 by the number of learning samples N is given.

Then, (a) learning of weak classifiers, calculation of (b) weighted error e _t, calculation of the weight of (c) weighting majority, (d) calculating data weights, the process consisting of the calculation of (e) termination threshold By repeatedly performing the number of times corresponding to the number K of weak classifiers to be generated, K (for example, several thousand) weak classifiers can be obtained.

In (a), the parameter of the t-th weak discriminator 21 _t is learned by Logistic Regression. The details of the weak classifier learning method will be described later.

The weighting error e _t calculated in (b) is an error when all learning samples are discriminated by the t-th weak discriminator 21 _t learned in (a), that is, the weak discriminator is a face image sample. When misclassified as a non-face, or conversely, when a non-face image sample is misclassified as a face, this is the total of the data weights D _{t, i} possessed by each misclassified learning sample. The weighting error _et is calculated using the following equation (15).

Subsequently, in (c), when the combiner 22 in the discriminator 5 performs a weighted majority decision on the output of each weak discriminator, the weight value α _t to be given to the weak discriminator 21 _t is obtained. The weight value α _t corresponds to the reliability of the weak classifier 21 _t and is calculated by the following equation (16).

Subsequently, in (d), a data weight D _{t + 1, i} indicating the difficulty of discrimination of each learning sample x ⁽ⁱ⁾ for the subsequent weak discriminator 21 _{t + 1} is represented by the following equation (17). as such, the weight alpha _t weighting majority, the label y ⁽ⁱ⁾ of each of the learning samples x ^(i), is calculated on the basis of the discrimination results for each training sample ^{x (i) h (x (} i)).

In the present embodiment, when performing the weighted majority decision by the adder 22, all without waiting for the calculation results of the weak discriminators _{21 t (21 1 ~21 K)} , is judged not to be the target object depending on the calculation middle value To stop the calculation. In (e), an abort threshold value for aborting the discrimination process is calculated. For example, among the labeled learning samples, the minimum value that can be taken by the weighted majority value of the discrimination result of the learning sample indicating the detection target can be set as the abort threshold. In the discrimination step, when the feature amounts are sequentially calculated by the weak discriminators of the window image, these are weighted and added as shown in the following equation (18), and the weighted majority value F is sequentially updated.

Each time a weak classifier outputs a feature value and updates the value of the weighted majority vote, it is compared with the abort threshold value, and the window image is not an object when the weighted majority vote value falls below the abort threshold value. By aborting the calculation, it is possible to speed up the discrimination process without using unnecessary calculations. K-th weak discriminator output fK termination threshold R _K of (x), of the learning samples _{x i (= x 1 ~x N} ), which is the object learning samples _{x j (= x 1 ~x J} ) Is the minimum value of the weighted majority vote, and is defined as in the following equation (19). In any case, the truncation threshold _RK is set to a maximum value that allows at least all positive learning samples to pass.

Each time the feature value calculated by the weak discriminator is multiplied by the reliability for the weak discriminator and added to update the weighted majority value, that is, the evaluation value s, the evaluation value s is compared with the truncation threshold value R _t. It is determined whether or not to continue the calculation of the estimated value of the discriminator. When the evaluation value s obtained by the weighted majority decision is equal to or less than the cutoff threshold R _t , even if the window image is cut out by scanning all regions of the input image and the scaled image obtained by reducing the input image, The probability of being an object is small and most are non-objects. Therefore, truncation of the calculation of the next subsequent weak discriminators when the evaluation value s falls below the abort threshold value R _t, by moving to the next window image processing, object dramatically reduce wasteful computing Detection can be speeded up. However, the calculation of the weighted majority vote is aborted in the middle, and the calculation of the abort threshold in (e) is not included in the original Adaboost algorithm.

Next, learning of the weak classifier (a) will be described in detail. The weak discriminator shown in the above equation (6) includes reference pixel positions {p ₁ , p ₂ , p ₃ , p ₄ }, filter coefficients {w ₁ , w ₂ , w ₃ , w ₄ }, and bias. consisting of parameters of w _5. Among these, the position of the reference pixel is searched M times (for example, several tens of thousands), and the filter coefficient and the bias are obtained for each reference pixel position by using Logistic Regression.

  FIG. 11 shows an algorithm for learning a weak classifier.

First, as a step 1 (a), when a set of filter reference pixel positions p _j is determined at random, a filter coefficient and a bias w _j that minimizes a cost function are obtained using Logistic Regression. Then, in step 1 (b), the weighted error e _m of the obtained 4-point filter or a weak classifier is calculated using the above equation (15) (where, m is an integer of 1 to M).

  In the set of reference pixel positions determined at random, the process of Step 1 of calculating the filter coefficient, the bias, and the weighting error by Logistic Regression (or linear SVM) is repeated M times. In step 2, a set of reference pixel positions where the weighting error is minimized, a filter coefficient, and a bias are adopted as parameters of the weak classifier.

When determining the parameter w _j (where j is an integer of 1 to 5) in step 1 (b), it is common to determine so as to minimize the weighting error expressed by the above equation (15). However, in the present embodiment, as shown in the following equations (20) to (22), the weighting cost function multiplied by the data weight D _{t, i} is minimized. Even if the cost function is not completely the same as the above equation (15), it should be understood that the Logistic Regi- sion normally functions as long as it is based on good discrimination.

  FIG. 12 shows, in the form of a flowchart, the processing procedure for learning a weak classifier using Adaboost that has been described so far.

  At the initial stage, since the difficulty level of discrimination of each learning sample is unknown, first, the data weight of each learning sample is initialized using the above equation (14) (step S1).

  Next, in a separately defined processing step S2, a weak classifier is learned with a weight (step S2).

  Then, for the learned weak classifier, parameters such as a weighting error, weighting weighting majority, and data weighting are updated (step S3). At that time, the abort threshold value of the weak classifier may be updated together.

  By repeating steps S2 and S3 K times (step S4), K weak classifiers can be learned. Then, the obtained linear sum of the K weak classifiers becomes the classifier 5 (step S5).

  Next, the procedure for learning one weak classifier in step S2 will be described.

  First, when a set of reference pixel positions is randomly determined (step S11), a filter coefficient and a bias are learned using logistic regression (or linear SVM) (step S12).

  Next, a weighting error is calculated for the weak classifier composed of the obtained filter coefficient and bias (step S13). If this weighting error is the smallest obtained so far, a set of reference pixel positions, The weak discriminator parameters including the filter coefficient and the bias are stored (step S14).

  The learning of the filter coefficient and the bias in the set of random reference pixel positions in steps S11 to S14 is performed M times (step S15). The parameter of the weak discriminator that minimizes the weighting error, that is, the cost function is set as a return value (step S16).

  FIG. 13 shows an example of a weak classifier, that is, a filter learned by the processing procedure shown in FIG. However, a multi-layered image (see FIG. 9) in which 2 × 2 layer and 4 × 4 layer blurred images are connected to the original input image of 1 × 1 layer was used as the learning sample. . Further, a four-point filter for front face detection is shown, and in the figure, filters selected from the first to the fifteenth are shown. In general, in boosting, the reliability of weak classifiers is high in the order of selection, and therefore the weight of the weighted majority vote becomes large.

  In FIG. 13, four squares indicate the positions of the filters, and white to black densities therein indicate the values of the filter coefficients. It can be seen that the first filter in the upper left is the filter most responsive to an image with dark eye position brightness and bright cheek position brightness.

Object detection:
Next, a method for discriminating an object from an input image in the object detection apparatus shown in FIG. 1 will be described. In the detection or discrimination step, the target object is detected from the input image using the discriminator 5 using the weak discriminator (filter) group generated as described above.

  FIG. 14 shows a processing procedure for detecting an object from an image using a weak classifier group in the form of a flowchart.

  First, the scaling unit 3 reduces the grayscale image given from the image input unit 2 at a certain rate (step S21), and outputs the scaled image to the scanning unit 4. A grayscale image may be input to the image input unit 2, or the input image may be converted into a grayscale image by the image input unit 2.

  Here, the timing for generating the scaled image is the time when the face detection of the entire area of the scaled image output previously is completed, and when the scaled image becomes smaller than the window image, the process proceeds to the processing of the input image of the next frame.

  Next, the scanning unit 4 scans the position of the search window vertically and horizontally on the scaled image, and outputs a window image (step S22).

Next, the discriminator 5 determines whether or not the window image output from the scanning unit 4 is an object. That is, a plurality of weak discriminators 21 _t composed of a four-point filter sequentially calculate the feature value f (x) for the window image (step S 23), and the combiner 22 sequentially weights and adds the weights. The update value of the majority value is calculated as the evaluation value s (step S24). However, when a new window image is input to the discriminator 5, first, the evaluation value s = 0 is initialized.

  Next, based on the evaluation value s, it is determined whether or not the window image is an object and whether or not the determination is terminated (step S25).

Next, the discriminator 5 determines whether or not the obtained evaluation value s is larger than the cutoff threshold R _t of the weak discriminator 21 _t (step S25). Here, when the evaluation value s is greater than the termination threshold R _t (Yes in step S25), and processing a predetermined number of times (= K times) of step S23~S24 only or repeated, i.e., the feature quantity by all the weak classifiers It is determined whether or not the calculation and weighted majority value update processing has been completed (step S26). If it has not been repeated a predetermined number of times yet (No in Step S26), the process returns to Step S23, and the calculation of the feature amount and the update process of the weighted majority value are repeated for the next weak discriminator 21 _{t + 1} .

On the other hand, when the processes in steps S23 to S24 are repeated a predetermined number of times (= K times) (Yes in step S26), the process proceeds to step S27. Further, when the evaluation value s becomes less than the abort threshold value R _t is abort the update process of the calculated value of the weighted majority decision feature quantity by the following stages of weak discriminators, one after step S27 is skipped and the next step S26 move on. Abort the calculation of the next stage and subsequent weak discriminators when the evaluation value s falls below the abort threshold value R _t, by moving to the next window image processing, a dramatically reduced by object detecting wasteful operations It is possible to increase the speed.

  In step S27, it is determined whether or not the object is included in the window image according to whether or not the obtained evaluation value s is greater than 0 (that is, whether the sign is positive or negative). When the object is included in the window image, the current window position is stored, and it is determined whether there is a next search window (step S27).

  When there is a next search window (Yes in step S27), the process returns to step S22, and when the scanning unit 4 cuts out the next search window, the processes of S23 to S26 are repeated.

  When the search window has been scanned for all the regions of the input image (No in step S27), the process proceeds to the next step S28 to determine whether or not the next scaling image is input from the scaling unit 3.

  When there is a next scaled image (Yes in step S28), the process returns to step S21, and the same processing is repeated for the next scaled image input from the scaling unit 3. The scaling process in step S21 ends when the scaled image becomes smaller than the window image.

  When processing of all the scaled images is completed for one input image (No in step S28), the process proceeds to overlap region deletion processing and is determined to be a target object in one input image. When overlapping areas overlap, the overlapping areas are removed. For this processing, first, it is determined whether or not there are areas overlapping each other (step S29). When there are a plurality of areas stored in step S26 and there are overlapping areas, the process proceeds to step S30. Then, two regions that overlap each other are extracted, and the region having a small evaluation value s is regarded as having low reliability, and the region having a large evaluation value s is selected from the two regions (step S29). Then, the processing from step S29 is repeated again. As a result, only one region having the highest evaluation value s is selected from the regions that are extracted multiple times. If two or more object areas do not overlap or there is no object area, the process for one input image is terminated, and the process proceeds to the next frame process.

  According to the object detection method of the present embodiment, the object is detected using a group of weak classifiers (filters) having a high degree of freedom in which each filter coefficient is an ideal real value, so that recognition performance is improved. .

  In addition, by adopting the truncation process in the process of updating the value of the weighted majority using the weak classifier group, it is possible to drastically reduce wasteful computation and speed up object detection. Even if the window image is cut out by scanning all regions of the input image and the scaled image obtained by reducing the input image, the censoring process is performed when it is considered that the probability of being an object among the window images is small. On the contrary, when a large number of objects to be detected are included, a threshold value may be provided so that the calculation of the window image that is clearly identified as the object is interrupted in the same way as the above-described threshold value. Good. Further, by scaling the input image by the scaling unit 3, a search window having an arbitrary size can be set and an object having an arbitrary size can be detected.

  FIG. 15 shows a state where a face area is detected as an object from an image photographed by a digital camera using a weak classifier group learned by Adaboost. By scaling the input image and scanning with a window image of a predetermined size, it is possible to detect suitably a plurality of face images scattered in the input image and varying in size due to perspective, etc. I want you to understand.

Filter variations:
Up to this point, the four-point filter (see FIG. 7) that calculates the feature amount by multiplying the luminance values of the four reference pixels by the filter coefficients has been described as the weak classifier. However, the gist of the present invention is that the number of reference pixels constituting the filter is not limited to four, and there are variations in the shape of the filter. For example, as shown in FIG. 16, a multipoint filter that refers to a total of 16 pixels of a rectangular 4 × 4 pixel can be considered. In the figure, there are three types of reference pixel sizes depending on the multi-layered image (described above). In the case of a four-point filter, it is only necessary to have two pieces of information of {start position q, step width s} to represent the positions of the four reference pixels (described above). On the other hand, in the case of a 4 × 4 pixel 16-point filter, it is sufficient to have two pieces of information: {start position q, step width s, window width w}, and the pixel position p _{i, j} is expressed by the following equation: It is expressed in

  Similarly to Patent Document 1, a two-point filter that refers to arbitrary two pixels is also conceivable. The pixel position may hold information for two points as it is. FIG. 17 shows a configuration example of the two-point filter. In the figure, there are three types of reference pixel sizes depending on the multi-layered image (described above). In the invention described in Patent Document 1, filter coefficients fixed to {+1, −1} are used (see the above equation (2)). On the other hand, in the present invention, the luminance value of each reference pixel is set. On the other hand, a high-degree-of-freedom filter, that is, a weak discriminator that multiplies an ideal real value filter coefficient, is learned, and the detection performance is improved by detecting an object using such a weak discriminator group.

  Various filters other than those described above can be considered, but the points are the following two points.

(1) Since the amount of calculation increases when the number of reference pixels is increased, the number of reference pixels is made as small as possible.
(2) The position information of the reference pixel can be held in as compact an expression as possible. This reduces the number of memory accesses that occur during execution, leading to higher speed.

  The learned discriminator consists of variables shown in the following table. By expressing and holding this in as small a data size as possible, the number of memory accesses that occur during execution is reduced, leading to higher speed. In the table, the number of bits necessary for holding is also illustrated.

To create a multi-layered image:
As described above, in the present embodiment, by performing image multi-layering (see FIG. 9), a wide range is referred to at a time in an original image of 1 × 1 layer. A processing module for processing the input image into a multi-layer can be disposed, for example, in the subsequent stage of the scanning unit 4 or incorporated in the scanning unit 4. Here, a method of creating a multi-layered image will be described.

  One method of multi-layering the original image is to create a low-resolution image of 10 × 10 pixels on the average of 2 × 2 pixels using 20 × 20 pixels of the input image, and then out of 10 × 10 pixels A low-resolution image of 5 × 5 pixels is created with an average of 2 × 2 pixels (an average of 4 × 4 pixels when viewed from an input image of 20 × 20 pixels), and 525 pixels including these three layers are combined as an input image. (See FIG. 18A). As another method, low-resolution images of 19 × 19, 17 × 17, and 13 × 13 are respectively generated by shifting the average filter of 2 × 2, 4 × 4, and 8 × 8 by one pixel. A method of combining all of them and using 1219 pixels as an input image can be cited (see FIG. 18B). In the example shown in FIG. 19, each pixel of the low-resolution image is the average of the square area of the original image. As a result, when a low-resolution pixel is selected, a plurality of original pixels are processed at a time, which increases efficiency. Comparison between areas of different sizes is also possible.

  When actual recognition processing is to be performed at high speed, the time for generating a multilayer image becomes a problem. If a multi-scale image is generated from an input image every time, the same pixel is averaged many times, and the efficiency is low. In order to eliminate processing redundancy, it is conceivable that the entire input screen is subjected to a multi-layer conversion process only once, and an image of a window portion to be actually recognized is cut out from each scale. Such a method is effective when there is a margin in memory.

  The inventors of the present invention propose a method for generating a multi-layer image in which the amount of memory used is suppressed as much as possible, and the redundancy of processing required for multi-layer conversion is eliminated while the number of memory copies is reduced.

  First, the data arrangement format is changed to a dot sequential method as shown in FIG. That is, the first four pixels of each scale (four layers in the example shown in the figure), followed by the next pixel taken from each layer. Therefore, if the size of the original window is 20 × 20 pixels, a window of 80 × 20 pixels is prepared. In order to generate this image, first four original 20 × 20 pixels are skipped and copied from the image area to this processing buffer.

  The lower layer pixels are then generated from the average of the four most recent upper layer pixels. The pixel of the second layer is generated from adjacent 2 × 2 pixels in the layer one level above, the third layer is generated from 2 × 2 pixels skipping one of the second layer, and the fourth layer is generated from the third layer Note that it generates from 2x2 skips.

  If this is left, the above-described processing occurs every time the window is moved. Therefore, an area of (win_width × (image_height−1) + image_width) × num_scale size as shown on the right side of FIGS. At the very beginning of processing the input image, four pixels of 20 × img_height at the left end of the input image are skipped as described above and copied to the strip buffer. Next, a multilayer image of 20 × image_height is created in a burst manner by the above-described procedure. As a result, a multi-slayer image corresponding to one vertical column of the input image is formed. Accordingly, recognition processing is performed while shifting the start address of the processing window from top to bottom.

  Since the dot sequence is used at this time, the positional relationship of the pixels in the window is always maintained as shown in FIG. 20 simply by moving the head address. Therefore, it is not necessary to newly copy or scale the input image while processing for one column.

  Next, a method for generating an image of the next column after the processing for one column is completed will be described. Since the window has moved only by one pixel (the number of pixels to be skipped when skip processing is being performed), most of the target pixels remain in the strip buffer. Therefore, copying is performed at an interval of win_width × numscale + 1 to win_width × num_scale in the strip for the newly required one column (or the skipped column). Then, a lower layer pixel corresponding to the filled pixel is generated by four averages of the pixels in the strip buffer.

  In the recognition process, the second pixel of the strip buffer is designated as the head, and similarly, the process for one column can be performed one after another.

  When processing the I-th column, fill the pixels for one column from the win_width × num_scale + i−1th pixel of the strip buffer, and specify the top of the processing buffer from the i−1th to process. Can do. As a result, the entire screen can be processed efficiently with a small buffer size.

  In the case of learning using boosting, a stamp classifier as shown in FIG. 22 is often used. In addition, a learning algorithm that takes into account the reliability of weak hypotheses WL such as GentleBoost and RealAdaBoost has been proposed using a regression stamp as shown in FIG.

  A weak discriminator for calculating the reliability generally uses a lookup table as shown in FIG. 24. The reliability is output in a bin by dividing feature quantities into equal parts. Yes. From the histogram shown in FIG. 24, the probability value of the positive / negative determination for each feature amount can be output more finely.

  However, in the case of an image recognition task using a pixel difference as disclosed in Patent Document 1, the difference value often has a distribution centered on 0 as in the histogram shown in FIG. This is because the closer pixels often take similar values. Therefore, if the feature amount space is evenly divided like the piece-wise function, a difference near 0 where information is most concentrated is overlooked. Focusing on this point, a multi-threshold regression stamp was proposed in which the division of the regression stamp that can freely set the break is increased by increasing the threshold value, and used as a weak discriminator of the pixel difference feature quantity.

  As described above, a pixel obtained by averaging a plurality of pixels is calculated and prepared in advance. Depending on the size of the area, prepare 3 types of images, such as averaged pixel of 2x2 area, averaged pixel of 4x4 area, and averaged pixel of 8x8 area. Together with the original image, it is called a multilayer image. Once the multi-layer image is calculated, it is possible to handle both fine patterns and a broad range of patterns when learning weak classifiers and detecting objects using weak classifiers. Can be reduced.

  The present invention has been described in detail above with reference to specific embodiments. However, it is obvious that those skilled in the art can make modifications and substitutions of the embodiment without departing from the gist of the present invention.

  The face / non-face discrimination problem handled by the present invention can be used as a “detection” technique for a face by scanning an image. Similarly, the present invention can also be applied to patterns other than the face, such as logo mark detection.

  Also, when the present invention handles a face, smile detection can be performed by further determining “smile / non-smile” for the face area after detecting the face. Of course, the present invention can be similarly applied to discrimination of facial expressions and states other than smiles.

  Alternatively, the present invention can also be applied to race discrimination and the like. Here, the race spans multiple classes such as “Yellow Race / White / Black”, but “Yellow Race / Non-yellow Race”, “White / Non-White”, “Black / Non-Black”, etc. This can be realized by converting the problem into a two-class problem. The same applies to other discriminations over three or more classes.

  The present invention can be used for detection, discrimination, and identification of various images, examples of which are given below.

(1) Logo mark detection (company logo, road signs, etc.)
(2) Facial expression discrimination such as smiles (3) Discrimination of eyes and mouth open / closed state (4) Gender discrimination (5) Adult / child discrimination (6) Race discrimination (7) Personal identification (whether it is a specific individual Is not)
(8) Determination of the presence or absence of glasses (9) Position detection of facial parts (face organs such as eyes, nose, mouth) (10) Character recognition (11) Car detection / model determination

  In short, the present invention has been disclosed in the form of exemplification, and the description of the present specification should not be interpreted in a limited manner. In order to determine the gist of the present invention, the claims should be taken into consideration.

FIG. 1 is a diagram schematically illustrating a functional configuration of the object detection apparatus. FIG. 2 is a diagram illustrating how the scaling unit 3 generates a reduced image. FIG. 3 is a diagram showing how the scanning unit 4 scans a window having a predetermined window size on the input image. FIG. 4 is a diagram showing the configuration of the discriminator 5. FIG. 5 is a diagram illustrating sample images of a face image and a non-face image. FIG. 6 is a diagram showing how the inter-pixel difference feature is obtained. FIG. 7 is a diagram illustrating a configuration example of a four-point filter. FIG. 8A is a diagram illustrating a state in which a linear four-point filter is arranged in an image that has been subjected to blurring processing by averaging luminance values for each 2 × 2 pixel block to obtain one pixel. FIG. 8B is a diagram illustrating a state in which a linear four-point filter is arranged on an image that has been subjected to blurring processing by averaging luminance values for each 4 × 4 pixel block to obtain one pixel. FIG. 9 is a diagram illustrating a state in which the 2 × 2 layer and 4 × 4 layer blurred images are connected to the original input image of the 1 × 1 layer to form a multi-layered image. FIG. 10 is a diagram showing boosting (Adaboost), which is a framework for learning. FIG. 11 is a diagram illustrating an algorithm for learning a weak classifier using Logistic Regression. FIG. 12 is a flowchart showing a processing procedure for learning a weak classifier by Adaboost. FIG. 13 is a diagram showing an example of a weak classifier, that is, a filter learned by the above-described method. FIG. 14 is a flowchart showing a processing procedure for detecting an object from an image using a weak classifier group learned by Adaboost. FIG. 15 is a diagram showing a state in which a face area is detected as an object from an image photographed by a digital camera using a weak classifier group learned by Adaboost. FIG. 16 is a diagram illustrating the configuration of a multipoint filter that refers to a total of 16 pixels of a rectangular 4 × 4 pixel. FIG. 17 is a diagram showing a configuration example of a two-point filter learned by applying the present invention. FIG. 18A is a diagram for explaining a method of creating a multilayer image. FIG. 18B is a diagram for describing a method of creating a multilayer image. FIG. 19 is a diagram illustrating a configuration example of a multilayer image. FIG. 20 is a diagram for explaining a method of changing the data arrangement format to the dot sequential method. FIG. 21A is a diagram for explaining processing of an input image using a processing buffer. FIG. 21B is a diagram for describing processing of an input image using a processing buffer. FIG. 21C is a diagram for describing processing of an input image using a processing buffer. FIG. 21D is a diagram for describing processing of an input image using a processing buffer. FIG. 21E is a diagram for describing processing of an input image using a processing buffer. FIG. 22 is a diagram illustrating a configuration example of the stamp classifier. FIG. 23 is a diagram illustrating a configuration example of a regression stamp. FIG. 24 is a diagram illustrating a state in which the reliability of outputting each bin by dividing the feature amount into equal parts is held in a table. FIG. 25 is a diagram illustrating an example of an image recognition procedure. FIG. 26 is a diagram for explaining the Sparse feature proposed by Huang et al.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Object detection apparatus 2 ... Image input part 3 ... Scaling part 4 ... Scanning part 5 ... Discriminator 6 ... Collective learning machine 21 ... Weak discriminator 22 ... Adder

Claims (23)

Object detection that detects whether a given image is an object using a learning sample consisting of multiple grayscale images with known correct answers (whether they are objects or non-objects) A collective learning device for collective learning of a plurality of weak classifiers used in
Each weak discriminator is a multi-point filter that calculates a feature value by calculating an inner product of a luminance value vector whose element is the luminance value of a reference pixel at the pixel position of the L point and a filter coefficient vector composed of an arbitrary real value ( However, L is an integer greater than or equal to 2),
For each weak discriminator, a learning means for learning a combination of reference pixel positions and an ideal real value used for a filter coefficient by boosting is provided.
A group learning device characterized by that. The learning means obtains an ideal real value to be used as a filter coefficient by reducing learning of a multipoint filter to a linear two-class discrimination problem.
The group learning apparatus according to claim 1. The learning means inputs a plurality of learning samples consisting of gray images in which two classes of objects or non-objects are classified, that is, labeled, into each weak classifier, and performs object and non-objects by boosting. Learn each feature amount,
The group learning apparatus according to claim 1. The input image (learning sample as well as the given image) consists of a predetermined window size,
Each of the plurality of weak classifiers arranges L reference pixels in a straight line on the window size, and holds each pixel position as two pieces of information of {start position q, step width s}.
The group learning apparatus according to claim 1. The original 1 × 1 layer input image is subjected to 2 × 2 blurring processing that averages the luminance value for each adjacent 2 × 2 pixel block to 1 pixel to generate a 2 × 2 layer image, and And a multi-layer image generation means for generating a 4 × 4 layer image by performing a 4 × 4 blurring process that averages the luminance value for each adjacent 2 × 2 pixel block of the original learning sample to 1 pixel. Prepared,
The learning means learns a weak classifier using a learning sample of each layer that is multi-layered by the multi-layer image generating means.
The group learning apparatus according to claim 1. The learning means determines a parameter of a multipoint filter including a combination of pixel positions of L points of reference pixels and a filter coefficient so as to minimize a discrimination error obtained by inputting a plurality of learning samples into a weak discriminator. To learn weak classifiers,
The group learning apparatus according to claim 1. The learning means determines a parameter of a multipoint filter including a combination of a pixel position of the L point of the reference pixel and a filter coefficient so as to minimize a discriminating error by using Logistic Regression or a support vector machine.
The group learning apparatus according to claim 6. Each learning sample i has a data weight D _{t, i} representing a weight reflecting the difficulty of discrimination for each weak discriminator t (where i is a serial number for identifying a learning sample, and t is a weak discriminator). Serial number),
The learning means includes
Initialization means for initializing data weights D _{1, i} of each learning sample for the weak classifier to be generated first;
Weak classifier learning means for learning the t-th weak classifier by using Logistic Regression or a support vector machine;
When a plurality of learning samples are input to the t-th weak classifier learned by the weak classifier learning means, a data error D _{t, i} of each misclassified learning sample is added to calculate a weighting error. A weighting error calculating means;
A weighted majority voting weight calculating means for calculating a weight α _t corresponding to the reliability of the t-th weak classifier, which is used when the weighted majority of the feature quantities calculated by each weak classifier at the time of detection of an object;
The data weight D _{t + 1, i} of each learning sample is calculated based on the weighting majority weight α _t , the label for correctly attaching each learning sample, and the discrimination result for each learning sample by the t-th weak discriminator. Data weight updating means to perform,
With
The weak classifier learning means, the weighting error calculation means, the weight calculation means, and the data weight update means repeat the process as many times as necessary to learn the required number of weak classifiers.
The group learning apparatus according to claim 6. The weak classifier learning means includes:
A weighting error obtained by adding the data weights D _{t, i} of misclassified learning samples when a combination of reference pixel positions of the multi-point filter t is randomly determined and a plurality of learning samples are input to the multi-point filter t The filter coefficient of the multi-point filter t that minimizes the cost function consisting of is determined using Logistic Regression or a support vector machine, and the weighting error of the multi-point filter t having the determined filter coefficient is obtained. Repeat the process several times,
A multi-point filter t composed of a combination of reference pixel positions and a filter coefficient that minimizes a weighting error in the above iterative processing is output as a t-th weak discriminator;
The group learning apparatus according to claim 8. At the time of detection of an object, a truncation process is introduced in which the given grayscale image is determined not to be an object according to a value in the middle of the calculation in which the feature quantity calculated by each weak classifier is weighted, and the calculation is terminated. And
A censoring threshold calculating means for calculating a censoring threshold based on a minimum value that can be taken by the weighted majority value of the discrimination result of the learning sample indicating the object;
When the processing by the weak discriminator learning unit, the weighting error calculation unit, the weight calculation unit, and the data weight update unit is repeatedly performed as many times as necessary, the processing by the truncation threshold value calculation unit is also performed, Calculate the censoring threshold for each weak classifier learned,
The group learning apparatus according to claim 8. Object detection that detects whether a given image is an object using a learning sample consisting of multiple grayscale images with known correct answers (whether they are objects or non-objects) A collective learning device for collective learning of a plurality of weak classifiers used in
Each weak discriminator is a multi-point filter that calculates a feature value by calculating an inner product of a luminance value vector whose element is the luminance value of a reference pixel at the pixel position of the L point and a filter coefficient vector composed of an arbitrary real value ( However, L is an integer of 2 or more, and each learning sample i has a data weight D _{t, i} that represents a weight reflecting the difficulty of discrimination for each weak discriminator t (where i identifies a learning sample) Serial number, t is a serial number that identifies the weak classifier),
An initialization step for initializing the data weights D _{1, i} of each learning sample for the weak classifier to be generated first;
A weak discriminator learning step of learning the t-th weak discriminator using Logistic Regression or a support vector machine;
When a plurality of learning samples are input to the t-th weak classifier learned in the weak classifier learning step, the weighting error is calculated by adding the data weights D _{t, i} of the misclassified learning samples. A weighting error calculation step;
A weighted majority voting weight calculating step for calculating a weight α _t corresponding to the reliability of the t-th weak classifier, which is used when weighting majority voting the feature amount calculated by each weak classifier at the time of detecting an object;
The data weight D _{t + 1, i} of each learning sample is calculated based on the weighting majority weight α _t , the label for correctly attaching each learning sample, and the discrimination result for each learning sample by the t-th weak discriminator. A data weight update step,
With
The weak classifier learning step, the weighting error calculation step, the weight calculation step, and the data weight update step are repeated as many times as necessary to learn the required number of weak classifiers.
A group learning method characterized by this. Object detection that detects whether a given image is an object using a learning sample consisting of multiple grayscale images with known correct answers (whether they are objects or non-objects) A computer program written in a computer-readable format so as to execute a process for collective learning of a plurality of weak classifiers used in the computer,
Each weak discriminator is a multi-point filter that calculates a feature value by calculating an inner product of a luminance value vector whose element is the luminance value of a reference pixel at the pixel position of the L point and a filter coefficient vector composed of an arbitrary real value ( However, L is an integer of 2 or more, and each learning sample i has a data weight D _{t, i} that represents a weight reflecting the difficulty of discrimination for each weak discriminator t (where i identifies a learning sample) Serial number, t is a serial number that identifies the weak classifier),
The computer program causes the computer to
Initialization means for initializing data weights D _{1, i} of each learning sample for the weak classifier to be generated first;
Weak classifier learning means for learning the t-th weak classifier by using Logistic Regression or a support vector machine;
When a plurality of learning samples are input to the t-th weak classifier learned by the weak classifier learning means, a data error D _{t, i} of each misclassified learning sample is added to calculate a weighting error. A weighting error calculating means;
A weighted majority voting weight calculating means for calculating a weight α _t corresponding to the reliability of the t-th weak classifier, which is used when the weighted majority of the feature quantities calculated by each weak classifier at the time of detection of an object;
The data weight D _{t + 1, i} of each learning sample is calculated based on the weighting majority weight α _t , the label for correctly attaching each learning sample, and the discrimination result for each learning sample by the t-th weak discriminator. Data weight updating means to perform,
Function as
The weak classifier learning means, the weighting error calculation means, the weight calculation means, and the data weight update means repeat the process as many times as necessary to learn the required number of weak classifiers.
A computer program characterized by the above. An object detection device for detecting whether a given grayscale image is an object,
It consists of a multi-point filter that calculates the feature value of the given grayscale image by calculating the inner product of a luminance value vector whose element is the luminance value of the reference pixel at the L pixel position and a filter coefficient vector consisting of an arbitrary real value. A plurality of weak classifiers (where L is an integer of 2 or more),
A discriminator for discriminating whether or not the given grayscale image is an object based on the feature amount calculated by at least one of the plurality of weak discriminators;
An object detection apparatus comprising: The discriminator calculates a weighted majority value obtained by multiplying the feature amount calculated by each weak discriminator by a weight representing the reliability of the corresponding weak discriminator and adding the weight, and based on the weighted majority value To determine whether the given grayscale image is an object.
The object detection apparatus according to claim 13. Each of the plurality of weak classifiers uses, as a filter coefficient, an ideal real value obtained by reducing learning of a multipoint filter to a linear two-class classification problem.
The object detection apparatus according to claim 13. Each of the plurality of weak classifiers inputs a plurality of learning samples composed of gray images that are classified, that is, labeled, as two classes of objects or non-objects, and each of the objects and non-objects by boosting Is a multi-point filter that has been learned in advance,
The object detection apparatus according to claim 13. The input image consists of a predetermined window size,
Each of the plurality of weak classifiers arranges L reference pixels in a straight line on the window size, and holds each pixel position as two pieces of information of {start position q, step width s}.
The object detection apparatus according to claim 13. The 2 × 2 blur processing is performed by averaging the luminance value of the 1 × 1 layer input image for each adjacent 2 × 2 pixel block to 1 pixel to generate a 2 × 2 layer image. A multi-layer image generating means for generating a 4 × 4 layer image by performing a 4 × 4 blurring process that averages the luminance value of each learning block of adjacent 2 × 2 pixel blocks to 1 pixel;
Each of the plurality of weak classifiers uses a learning sample of each layer that has been multi-layered, and a multi-point filter for any one of the layers has been learned in advance. Calculating a feature amount for a corresponding layer obtained by multi-layering a grayscale image by the multi-layer image generation unit;
The object detection apparatus according to claim 13. Each of the plurality of weak classifiers determines a parameter of a multipoint filter including a combination of pixel positions of L points of reference pixels and a filter coefficient so as to minimize a discrimination error obtained by inputting a plurality of learning samples. To be learned in advance,
The object detection apparatus according to claim 13. The parameters of the multi-point filter are determined so as to minimize the discrimination error of the weak discriminator using Logistic Regression or a support vector machine.
The object detection device according to claim 19. For each weak classifier, the censoring threshold is learned based on the minimum value that can be taken by the weighted majority value of the discrimination result of the learning sample indicating the object,
The discriminator determines that the given gray-scale image is not an object according to a comparison result between a halfway calculation value for weighted majority determination of the feature amount calculated by each weak discriminator, and the calculation is performed. abort,
The object detection apparatus according to claim 13. An object detection device for detecting whether a given grayscale image is an object,
A plurality of weak discrimination steps for calculating feature values of the given grayscale image by calculating an inner product of a luminance value vector whose element is the luminance value of the reference pixel at the L pixel position and a filter coefficient vector consisting of an arbitrary real value (Where L is an integer of 2 or more),
A determination step of determining whether or not the given grayscale image is an object based on the feature amount calculated by at least one or more of the plurality of weak determination steps;
An object detection apparatus comprising: A computer program written in a computer-readable format so as to execute on a computer a process for detecting whether or not a given grayscale image is an object, the computer comprising:
It consists of a multi-point filter that calculates the feature value of the given grayscale image by calculating the inner product of a luminance value vector whose element is the luminance value of the reference pixel at the L pixel position and a filter coefficient vector consisting of an arbitrary real value. A plurality of weak discrimination means (where L is an integer of 2 or more),
Discriminating means for discriminating whether or not the given grayscale image is an object based on the feature amount calculated by at least one of the plurality of weak discriminating means;
Computer program to function as