JP2014199533A - Image identification device, image identification method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To identify an image at a high speed while suppressing the storage capacity.SOLUTION: A learning section 201 can reduce the dictionary size by storing, in a dictionary, a lookup table corresponding to a selected feature category and the identification contribution of a non-selected feature category that is used as a correction value. A selected category feature vector acquisition section 222 can reduce the referring/writing processing of a memory corresponding to the resolution of an input image, by preventing the acquisition of the feature vector corresponding to the feature category that is not selected by the learning section 201.

Description

  The present invention particularly relates to an image identification device, an image identification method, and a program suitable for identifying an object based on a dictionary.

  To identify whether an image is an object class, obtain a feature vector from the image, create a dictionary using machine learning, and determine whether an image is an object based on the dictionary. Many methods of identifying are taken.

  Non-Patent Document 1 discloses a technique for speeding up the process by using the Histogram Intersection Knel and creating a lookup table as a technique for identifying whether or not an image is an object class. Non-Patent Document 2 discloses a kernel function that can create a lookup table in the same manner as the method described in Non-Patent Document 1. Further, the Support Vector Data Description described in Non-Patent Document 3 is a learning machine to which the creation of the Histogram Inter- ference Kernel and the lookup table described in Non-Patent Document 1 can be applied.

  Further, in order to identify which of a plurality of object classes an image is, a technique of preparing a plurality of feature categories for describing the characteristics of each object class is used. Non-Patent Documents 4 to 6 disclose HOG features, LBP features, and Bag of visual words features, which are representative feature categories, respectively.

  Further, Patent Document 1 discloses a method of obtaining a feature vector from a plurality of feature categories, obtaining a discriminator by machine learning, preparing a plurality of kernel functions, and obtaining an optimum weight for each kernel function. ing. According to the method described in Patent Document 1, when the characteristic categories that are different in the characteristics of a plurality of object classes and the effective feature categories are different, the identification is performed by learning the appropriate weight according to the effectiveness of the kernel. It is described to improve the rate.

US Patent Application Publication No. 2009/0099986

Subhran Maji, Alexander C. et al. Berg, Jitendra Malik, “Classification using Intersection Kernell Support Vector Machines is Efficient”, Computer Vision and Pattern Recognition, 2008 Andrew Vedaldi, Andrew Zisserman, Efficient Additive Kernels, Via Expert Feature Map, IEEE Transactions on Pattern Analysis and Mass Analysis and Pattern Analysis. D. M.M. J. et al. Tax and R.C. P. W. Duin. Support vector data description. Machine Learning, 2002 Dalal, N.A. , Triggs, B, “Histograms of orientated gradients for human detection”, Computer Vision and Pattern Recognition, 2005 Timo Ojala, Matti Pietikainen, Multiresolution Gray-Scale and Rotation Invariant Texture Classification 2 Local Binary Strings, IEEE TRANSITION L. Fei-Fei and P.M. Perona (2005). “A Bayesian Hierarchical Model for Learning Natural Scene Categories”. Proc. of IEEE Computer Vision and Pattern Recognition. pp. 524-531

  In general, when an image is identified, it is desirable that it is high speed and storage capacity can be suppressed. When creating a dictionary for identification using a plurality of different feature categories, the known method has a storage capacity for storing the feature vector dictionary according to the number of feature categories regardless of the validity of the feature categories. There is a problem that calculation time is required. For example, in the method described in Patent Document 1, since the feature vectors of all the feature categories to be used are computed, the feature vector computation time and the storage capacity for storing the dictionary are required according to the number of feature categories.

  An object of the present invention is to make it possible to identify an image at a high speed and with a small storage capacity.

  The image identification device of the present invention includes a first acquisition unit that acquires a feature vector indicating a feature of a learning sample image, and a feature vector of each learning category acquired from the feature vector of the learning sample image acquired by the first acquisition unit. A calculation unit for obtaining an identification contribution, a selection unit for selecting a feature category whose identification contribution obtained by the calculation unit is equal to or greater than a predetermined value, and a dictionary including information on the feature category selected by the selection unit A dictionary creating means, a second obtaining means for obtaining a feature vector indicating a feature of the input image, a feature vector of the input image obtained by the second obtaining means, and a dictionary created by the dictionary creating means Determination means for determining whether the object of the input image is similar to the object of the learning sample image based on the dictionary. Said second acquisition means, characterized by the feature vector corresponding to the feature category is not selected by said selecting means so as not obtained from the input image.

  According to the present invention, it is possible to speed up the process of acquiring the feature vector and identifying the input image, and to identify the image while suppressing the storage capacity of the dictionary.

It is a block diagram which shows the hardware structural example of the image identification device which concerns on embodiment of this invention. It is a block diagram which shows the function structural example of the image identification device which concerns on embodiment of this invention. It is a figure for demonstrating the method of acquiring and connecting a feature vector. In an embodiment, it is a flow chart which shows an example of a processing procedure in a learning part. It is a figure which shows the information group referred in the process of creating the dictionary A from the object A. FIG. It is a figure which shows the information group referred in the process of creating the dictionary B from the object B different from the object A. It is a figure for demonstrating the outline of the look-up table which speeds up the calculation of a discriminator. In an embodiment, it is a flow chart which shows an example of a processing procedure in an identification part. It is a figure which shows an example of the selection result of the feature category about the dictionary A corresponding to the object A, the correction value of the feature category which was not selected, and the lookup table with respect to the selected feature category. FIG. 9 is a flowchart showing an example of a detailed processing procedure for creating a feature vector for a selected feature category in step S801 in FIG.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
(First embodiment)
FIG. 1 is a block diagram illustrating a hardware configuration example of an image identification device 100 according to the present embodiment.
In FIG. 1, an image sensor 101 is constituted by a CCD, a MOS, or the like, and is an image pickup means for converting a subject image from light to an electrical signal. The signal processing circuit 102 processes a time series signal related to the subject image obtained from the image sensor 101 and converts it into a digital signal.

  The CPU 103 controls the entire image identification apparatus 100 by executing a control program stored in the ROM 104. The ROM 104 stores a control program executed by the CPU 103 and various parameter data. The control program is executed by the CPU 103, thereby causing the apparatus to function as various means for executing each process shown in a flowchart described below. The RAM 105 stores images and various information. The RAM 105 functions as a work area for the CPU 103 and a temporary data saving area. The display 106 is composed of, for example, an LCD or a CRT.

  In the present embodiment, processing corresponding to each step of the flowchart described later is realized by software using the CPU 103, but part or all of the processing is realized by hardware such as an electronic circuit. It doesn't matter. Further, the image identification apparatus of the present embodiment may be realized using a general-purpose personal computer (PC) without the image sensor 101 and the signal processing circuit 102, or may be realized as an image identification dedicated apparatus. I do not care. Further, software (program) acquired via a network or various storage media may be executed by a processing device (CPU, processor) such as a PC.

FIG. 2 is a block diagram illustrating a functional configuration example of the image identification device 100 according to the present embodiment.
In the image identification device 100 of the present embodiment, a plurality of learning sample images indicating the same object class are given in advance. In addition, the image identification device 100 includes a learning unit 201 that creates a dictionary of classifiers from the learning sample image group, and an identification unit 202 that identifies whether an image based on the dictionary is the same as the object class. Yes.

  In the image identification device 100 according to the present embodiment, the learning unit 201 reduces the size of the dictionary, and the identification unit 202 speeds up the identification of the object class. Note that the learning unit 201 and the identification unit 202 operate on the hardware configuration illustrated in FIG. 1, but may operate in different environments. For example, the learning unit 201 may operate on a PC via a network, and the apparatus illustrated in FIG. 1 may acquire the dictionary and operate the identification unit 202.

  The learning unit 201 creates a dictionary that stores an information group for identifying the object class from the learning sample image. The learning unit 201 includes a feature vector acquisition unit 211, a first feature vector normalization unit 212, a classifier learning unit 213, a lookup table creation unit 214, and an identification contribution degree calculation unit 215. Furthermore, a feature category selection unit 216, a feature category correction value calculation unit 217, and a dictionary creation unit 218 are provided. In the present embodiment, the size of the dictionary is reduced by obtaining the identification contribution for each feature category using the learning result obtained by acquiring the feature vector, and storing only the feature category contributing to the identification in the dictionary.

  The feature vector acquisition unit (first acquisition unit) 211 acquires a feature vector from each learning sample image. The feature vector is a histogram feature indicating a feature of an image, for example, a light and dark pattern or an edge pattern as a vector. This feature vector is generated by connecting features obtained from different feature categories, and each dimension belongs to one of the feature categories.

  FIG. 3 is a diagram for explaining a method of acquiring and connecting feature vectors. First, feature vectors corresponding to all feature categories are obtained for the learning sample image. Next, these feature vectors are connected. In the present embodiment, the feature categories are the HOG feature, the LBP feature, and the Bag of visual words (BoVW) feature described in Non-Patent Documents 4 to 6, respectively, but are not limited to these features. For example, feature vectors may be acquired from individual grids obtained by dividing an image into grids, and the X and Y coordinates of the acquired grids may be included in the feature categories, or different features may be introduced and new feature categories assigned. May be.

  The first feature vector normalization unit 212 normalizes the feature vector acquired by the feature vector acquisition unit 211 so as to be within a predetermined range. The normalization method is not particularly limited as long as it falls within this range.

  The classifier learning unit 213 learns a classifier from the feature vector group normalized by the first feature vector normalization unit 212. This discriminator is created by a learning machine based on a kernel function and a support vector described in Non-Patent Document 1 and Non-Patent Document 3. In other words, this method uses a kernel function instead of the inner product in the learning process, leaves a valid support vector group from the given feature vector group, and estimates the weight for each remaining support vector. It is a technique to do. In addition, this learning machine can generate a lookup table described later. As shown in Non-Patent Document 2, a lookup table can be generated by an identifier that performs identification based on a weighted sum of kernel inner products of feature vectors and support vectors. It is a function that can be expressed in the form of linear addition of operations.

The lookup table creation unit 214 creates a lookup table that stores approximate values of the computations of the support vectors learned by the discriminator learning unit 213 and their weights in order to speed up computations in the discrimination unit 202. In FIG. 5B, an example of a lookup table is shown in the rows from LUT (r = 0) to LUT (r = R _max ). In FIG. 5B, D means a dimension and r means a section. For example, the identification value of the section r = 1 of the dimension d = 1 is stored in the entry 516.

  Based on the lookup table created by the lookup table creation unit 214, the identification contribution degree calculation unit 215 obtains the degree to which each feature category contributes to image identification. The feature category selection unit 216 determines whether or not the identification contribution is a predetermined value or more based on the identification contribution for each feature category obtained by the identification contribution calculation unit 215, and selects the feature category used for identification To do.

  The feature category correction value calculation unit 217 sets a parameter that approximates the value that can be predicted to be added to the identification value by the identification value calculation unit 224 if selected for the feature category not selected by the feature category selection unit 216. Ask. The dictionary creation unit 218 collects a group of information necessary for image identification in the identification unit 202 as a dictionary and stores it in the RAM 105. Here, for the feature categories not selected by the feature category selection unit 216, the correction value obtained by the feature category correction value calculation unit 217 is stored instead of the lookup table, and the dictionary is made smaller.

  The identification unit 202 identifies whether an image is the object class based on the dictionary created by the learning unit 201 from the learning sample image group. The identification unit 202 includes a dictionary reading unit 221, a selected category feature vector acquisition unit 222, a second feature vector normalization unit 223, an identification value calculation unit 224, and an identification value correction unit 225. In this embodiment, the amount of calculation is reduced by acquiring feature vectors only for feature categories stored in the dictionary.

  The dictionary reading unit 221 reads the dictionary created by the dictionary creating unit 218 from the RAM 105. The selected category feature vector acquisition unit (second acquisition unit) 222 selects a feature vector acquisition unit 211 from the feature category selected by the feature category selection unit 216 with respect to an input image that identifies whether the object class is an object class. Similarly, a feature vector is acquired. Here, in many cases, pre-processing is required to acquire a feature vector, but pre-processing as well as acquisition of a feature vector is not performed for an unselected feature category.

  The second feature vector normalization unit 223 performs normalization equivalent to the first feature vector normalization unit 212 for the feature vector acquired by the selected category feature vector acquisition unit 222. The identification value calculation unit 224 calculates an identification value for identifying whether or not the feature vector is the object class with respect to the feature vector normalized by the second feature vector normalization unit 223. At this time, the identification value calculation unit 224 uses the lookup table created by the lookup table creation unit 214 and stored in the dictionary by the dictionary creation unit 218. According to the description in Non-Patent Document 1, the calculation time for identification is a constant time with respect to the number of support vectors, and is proportional only to the number of dimensions. However, a storage capacity proportional to the number of dimensions and the number of domain divisions is required.

  The identification value correction unit 225 corrects the identification value obtained by the identification value calculation unit 224 and identifies whether or not it is an object class. Specifically, after adding an approximate value of the weighted sum for the feature category not selected by the feature category selection unit 216 to the identification value, it is determined whether or not the identification value exceeds a predetermined threshold value. judge.

  Hereinafter, the operation of the learning unit 201 when a learning sample image group A indicating a certain object A and a learning sample image group B indicating an object B different from the object A are given will be described. FIG. 4 is a flowchart illustrating an example of a processing procedure in the learning unit 201. The flowchart shown in FIG. 4 shows a procedure for inputting a learning sample image group of an object for learning a classifier and outputting a dictionary of the classifier. FIG. 5 shows an information group referred to in the process of creating the dictionary A from the object A, and FIG. 6 shows an information group referred to in the process of creating the dictionary B from the object B.

  First, in step S401, the feature vector acquisition unit 211 acquires feature vectors for all the features included in the feature category for all the images included in the learning sample image group. Next, in step S402, the first feature vector normalization unit 212 normalizes the feature vector.

FIG. 5A shows an example of a feature vector acquired from the learning sample image group A. The feature categories are a slope direction histogram (HOG) feature, a local binary pattern (LBP) feature, and a Bag of visual words (BOVW) feature described in Non-Patent Documents 4 to 6. Each feature category is composed of vectors having dimensions of D0, D1,..., The total number of dimensions through all feature categories is D _max, and the value of each dimension of the feature vector is 0 by normalization. 1, it is assumed that the fall in the range of ··· V _max.

  A normalized feature vector group obtained from each of the learning sample image groups A is shown in each row of FIG. For example, the value 501 is a feature obtained from the image 2 by the HOG feature, and the value is 6. It is assumed that different feature vectors have been acquired for the learning sample image group B by the same means.

Next, in step S403, the classifier learning unit 213 learns a recognizer based on the feature vector, and the lookup table creation unit 214 creates the lookup table. For the learning sample image group A, a lookup table generated from the feature vector shown in FIG. 5A is displayed in each row of LUT (r = 0) to LUT (r = R _max ) in FIG. It is shown.

  As for the discriminator, it is assumed that “Histogram Intersection Kernel” which is a technique described in Non-Patent Document 1 is applied to “Support Vector Data Description” described in Non-Patent Document 3. That is, first, assuming a space having “histogram intersection kernel” as an inner product of vectors, a hypersphere having a minimum volume including as many feature vector groups as possible included in the class to be identified is obtained. Next, only the feature vectors necessary for the definition of the hypersphere are left as support vectors, the weights are calculated for the remaining individual support vectors, and finally, the classifier can determine that the feature vectors in the class to be identified are included. Find the threshold.

  This discriminator is a weighted sum of the feature vector to be identified and the inner product of the histogram intersection kernel of the support vector. That is, the identification value is obtained by the following equation (1). A discriminator is represented by whether or not this value exceeds the obtained threshold value.

In Expression (1), d represents a dimension, and SV represents an index of a support vector. X _d ^SV represents the value of the dimension d of the support vector SV, and x _d represents the value of the dimension d of the feature vector to be identified. Furthermore, α _SV represents the weight of the support vector SV.

  FIG. 7 is a diagram for explaining an outline of a lookup table for speeding up the operation of the discriminator represented by Expression (1). Details of the lookup table and a method for creating the lookup table are described in Non-Patent Document 1. It should be noted that the above formula (1) corresponds to the area of the blurred portion in FIG. 7A, and that the formula (1) can be expressed in the form of linear addition of operations of individual dimensions. The identification value of the dimension d is calculated by the following formula (2) and corresponds to the area under the curve shown in FIG.

Further, when Expression (2) is divided into cases where x _d is a boundary, the following Expression (3) is obtained.

When the value of the equation (3) is represented by x _d on the horizontal axis, it is as shown in FIG. Since the set of {SV | x _d ^SV <x _d } is obtained in logarithmic time by sorting x _d ^SV , equation (3) can also be calculated in logarithmic time.

Further, the calculation of equation (3) is made into a constant time. The value of the expression (3) is obtained for the representative value r _d of each section r obtained by dividing the domain of x _d ^SV and x _d , and the value is stored in the lookup table. When Expression (3) is obtained for the representative value r _d , the value LUT (r, d) corresponding to the dimension d and the section r of the lookup table is calculated by the following Expression (4).

Here, r is represents the index of the section feature quantity vector identifying fall, r _d denotes a representative value in the interval r which is the feature quantity vector that identifies fall. FIG. 7D shows the relationship between the identification value and the representative value r _d . However, the division of the definition area and the derivation of the representative value thereof are arbitrary. For example, the domain may be divided at equal intervals. In the present embodiment, the range from 0 V _max shall R _max divided, the index indicating the predetermined section r, the median of the interval the representative value r _d, i.e.,
V _max * (r + 0.5) / R _max
And

As shown in Expression (4), the value LUT (r, d) in a certain dimension d is monotonically increasing in the r direction. Therefore, the maximum value in each dimension of the lookup table obtained in step S404 is equal to the value when the section r takes the maximum value _Rmax .

Next, in step S404, the identification contribution degree calculation unit 215 obtains an identification contribution degree for each feature category from the lookup table created in step S403. Identification contribution of each feature category mean degree of contribution to the identification of the image, identification contribution takes values from 0 V _max, and thus it is close to the object larger. In the present embodiment, for each feature category, the maximum value of the lookup table of each dimension included in the feature category is obtained, and the average is used as the identification contribution. The maximum value of the lookup table means the maximum value that the feature category contributes to identification when determining that an unknown sample belongs to the object class, and the average means the contribution per dimension, that is, the dictionary size described later To do.

  For example, FIG. 5B shows the LUT maximum value 512 as the maximum value for each dimension of the lookup table 511 corresponding to the feature category BoVW generated from the learning sample image group A. In FIG. 5C, the identification contribution 522 is shown as an average value of the LUT maximum value 512. Similarly, the LUT maximum value 513 as the maximum value of the feature category HOG and the identification contribution 523 which is the average value thereof are shown. 6A shows the lookup table generated from the learning sample image group B and the maximum value obtained, and FIG. 6B shows the identification contribution for each feature category. ing.

  Next, in step S405, the feature category selection unit 216 selects a feature category used for recognition based on the identification contribution obtained in step S404. In this embodiment, only features having an identification contribution greater than or equal to a predetermined threshold are selected. For example, when the threshold value of the identification contribution is set to 30, the BoVW feature and the LBP feature are selected for the feature category of the object A as shown in the selection column of FIG. Similarly, for the feature category of the object B, the HOG feature and the LBP feature are selected as shown in the selection column of FIG.

  Next, in step S <b> 406, the feature category correction value calculation unit 217 obtains correction values for feature categories that have not been selected. The correction value is a value that approximates the weighted sum of the kernel inner product of the unknown feature vector and the support vector for the feature category that has not been selected. In the present embodiment, the maximum value of the lookup table obtained as the identification contribution is used as the correction value for the feature category. Setting the maximum value in the lookup table as a correction value means setting a correction value as a value that can be easily identified as belonging to the object class in the omitted feature category, and is intended to minimize false negatives in identification. ing. As shown in the correction value column of FIG. 5C, the correction value 521 of the HOG feature of the object A and the correction value 621 corresponding to the BoVW feature of the object B as shown in the correction value column of FIG. Is required.

  Next, in step S407, the dictionary creation unit 218 obtains the selection result for each feature category obtained in step S405, the lookup table corresponding to the selected feature category, and the correction value corresponding to the feature category that has not been selected. Is stored in the dictionary. In this embodiment, the threshold value of the identification value obtained in step S403 is also stored in the dictionary.

  For example, as shown in FIGS. 5B and 5C, for the dictionary generated from the feature A, a selection result 524, lookup tables 511 and 514, and a correction value 525 are stored. Note that the lookup tables 511 and 514 correspond to the selected BoVW feature and LBP feature. Then, the lookup table 515 corresponding to the feature category that has not been selected, that is, the HOG feature, is not stored in the dictionary, and the correction value 525 is stored instead.

  Similarly, the dictionary generated from the feature B is a lookup table corresponding to the selection result 622 for each feature category, the selected HOG feature, and the LBP feature, as shown in FIGS. 6 (A) and 6 (B). 611 and 612 are included. For the BoVW feature that has not been selected, only the corresponding correction value 623 is stored in the dictionary, and the lookup table 613 is not stored in the dictionary.

  As described above, according to the present embodiment, the learning unit 201 reduces the dictionary size by not storing the lookup table corresponding to the feature category not selected in Step S405 in the dictionary. Further, as shown in the step S405, when selecting a feature category, an appropriate different feature category is selected according to the feature of the object so that the selection result 524 of the feature A and the selection result 622 of the feature B are different. The Furthermore, in the present embodiment, after selecting a feature category, the identification contribution is used as a correction value for a feature category that has not been selected, and learning that refers to all feature vectors again is not performed. Thereby, it is faster than the method of re-learning using the selected feature category.

  FIG. 8 is a flowchart illustrating an example of an operation procedure of the identification unit 202 of the present embodiment. In the process of the flowchart shown in FIG. 8, the dictionary and the input image created by the learning unit 201 are input, and the result of whether or not the input image is similar to the learning sample image group input to the learning unit 201 is output. As described above, the dictionary stores whether or not each feature category has been selected, a lookup table for the selected feature category, and correction value information for the unselected feature categories. FIG. 9A shows the selection result of the feature category for the dictionary A corresponding to the object A and the correction value of the feature category that has not been selected. FIG. 9B shows a lookup table for the selected feature category.

  First, in step S801, the selected category feature vector acquisition unit 222 creates a feature vector for the selected feature category. Details of this process will be described later.

  Next, in step S802, the second feature vector normalization unit 223 normalizes the feature vector obtained in step S801. This process is equivalent to the process of step S402 described above. As described above, by the processing in step S801 and step S802, for example, a feature vector shown in the feature vector row of the input image in FIG. 9C is obtained for the input image.

  Next, in step S803, the identification value calculation unit 224 calculates the identification value based on the feature vector normalized in step S802. This calculation method of the identification value corresponds to the above-described recognizer and lookup table creation method in step S403. For identification of a certain feature vector x using a lookup table, the following formula (5) is used in which the formula (2) in formula (1) described in step S403 is replaced with formula (4).

Here, r _xd represents a section where x _d falls. That is, for each dimension d constituting the feature vector, an interval r _d is obtained from the value x _{d of the} feature vector, a lookup table is referenced based on the dimension d and the interval r _d , and a value LUT obtained from the lookup table Find the sum through the dimensions of (r _d , d). The interval r _xd is the integer part of the quotient obtained by dividing x _d by V _max / R _max . V _max is the maximum normalization value of the feature vector values defined in step S402 and step S802. R _max is the number of index divisions defined in step S403.

For feature vector indicating a row of the feature vector of the input image of FIG. 9 (C), max = 100 , R max = 10 and the time interval corresponding to the value 931 corresponding to d = 1 r _d is 12 ÷ ( 100 ÷ 10) ≈1. That is, it corresponds to the entry 921 in the lookup table shown in FIG. Similarly, the value 932 corresponding to d = 2 is 55 ÷ (100 ÷ 10) ≈5, and corresponds to the entry 922 in the lookup table. The result of referring to the lookup table for each dimension of the feature vector of the input image is shown in the LUT reference value section of FIG. The sum of the rows of the LUT reference values in FIG. 9C is the identification value obtained in this step. The obtained sum is shown in the item of the identification value of the selected feature category in FIG.

  The calculation amount in step S803 conventionally takes a calculation time depending on the number of dimensions and the number of support vectors, but in the expression (5), the use of a lookup table increases the calculation amount to a degree proportional to the number of dimensions. Is suppressed. However, if only conventional feature selection or dimension reduction is performed without any special contrivance, the effect of reducing the amount of calculation is such that the number of references to the lookup table can be reduced.

  Next, in step S804, the identification value correction unit 225 adds the correction value of the category not selected to the identification value. As shown in FIG. 9A, the category not selected in the case of the dictionary A is HOG, and the corresponding correction value 911 is added to the identification value. When there are a plurality of categories not selected, the correction values are added for all the categories. As a result of the addition, a value 941 shown in FIG. 9D is obtained.

  Next, in step S805, the identification value correction unit 225 determines whether or not the obtained identification value is greater than or equal to a certain threshold value, and determines whether or not the image is similar to an object according to the determination result. . In the present embodiment in which the technique described in Non-Patent Document 3 is applied as the discriminator in step S403, this threshold is determined at the time of learning in step S403 and is included in the dictionary. For example, if the threshold value obtained in step S403 is 100, the value 941 is larger than the threshold value, so it is determined that the input image is similar to the object.

  FIG. 10 is a flowchart illustrating an example of a detailed processing procedure for creating a feature vector for a selected feature category by the selected category feature vector acquisition unit 222 in step S801. In the process of this flowchart, a dictionary storing whether or not each feature category has been selected is input, and a feature vector is output. Note that the selection result column in FIG. 9A indicates whether or not each feature category has been selected.

  In step S1001, focusing on one feature category, steps S1002 to S1008 are repeated for each selected feature category.

  First, in step S1002, it is determined to which category the feature category focused in step S1001 belongs. As a result of the determination, if the feature category is BoVW, the process proceeds to step S1003, if it is HOG, the process proceeds to step S1005, and if it is LBP, the process proceeds to step S1007.

  In steps S1003 to S1004, the BoVW feature is extracted. First, local feature points are obtained as preprocessing in step S1003, and BoVW features are obtained from the local feature points obtained in step S1004. After acquiring the BoVW feature, if a feature category that is not focused remains, the process proceeds to step S1001.

  In steps S1005 to S1006, HOG features are extracted. First, in step S1005, a differential image is obtained, and in step S1006, the input image is divided into a grid, and HOG features are extracted for each. If a feature category that is not focused remains, the process proceeds to step S1001.

  In steps S1007 to S1008, LBP features are extracted. First, in step S1007, an integral image of brightness is obtained, and in step S1008, the input image is divided into a grid, and LBP features are extracted for each. If a feature category that is not focused remains, the process proceeds to step S1001.

  When the BoVW feature and the LBP feature are selected as shown in FIG. 9A, steps S1003 to S1004 and steps S1007 to S1008 corresponding to these are executed. In this case, Steps S1005 to S1006 corresponding to unselected HOG features are not executed.

  As described above, when the processing of steps S1002 to S1008 is repeated and features are obtained for all feature categories, the obtained features are combined and output as feature vectors in step S1009. For example, when the dictionary shown in FIG. 9B is followed, a BoVW feature and an LBP feature are obtained, and the result of the combination is shown in the feature vector row of the input image in FIG. 9C.

  As described above, according to the present embodiment, the selected category feature vector acquisition unit 222 is configured not to acquire feature vectors corresponding to feature categories not selected by the learning unit 201. As a result, the calculation amount of the selected category feature vector acquisition unit 222 can be made smaller than when the feature vectors of all feature categories are acquired. Also, pre-processing is required for acquiring many image features, at least any of BoVW, HOG, and LBP features, and the processing cost depends on the resolution for processing the input image. That is, by not selecting a certain feature category, the memory reference / write processing is reduced by the resolution of the input image. The amount of calculation reduction due to not selecting a certain feature category depends on the resolution of the input image, and is much less than the reduction in the number of lookup table references by feature selection as described in this embodiment. An effect is obtained.

(Second Embodiment)
Among the processes of step S403 described in the first embodiment, the process performed by the classifier learning unit 213 applies a plurality of methods if the learning machine is based on a support vector and the kernel function is an additive kernel. Is possible.

  For example, when a feature vector that does not belong to the object class can be prepared in addition to the feature vector that belongs to the object class, the one-class SVM may be applied as a learning machine. In addition to the feature vector belonging to the object class, when a large number of feature vectors not belonging to the object class can be prepared, a method based on the Support Vector Machine as described in Non-Patent Document 1 may be used. In these cases, in general, higher generalization performance can be expected in a subspace where a similar feature vector is not given at the time of learning than when SVDD is used.

  In addition, as additive kernel, Non-Patent Document 2 shows an RBF kernel and a GRBF kernel, and these can also be applied. In the case of these kernels, the inner product of feature vectors becomes more nonlinear than the Histogram Intersection, and higher discrimination performance can be expected in a partial space where positive examples and negative examples are complicated in the feature vector space.

  Note that, regardless of which combination is selected, the subsequent lookup table creation procedure is the same as in the first embodiment. However, when the Histogram Intersection Kernel is not used, it is necessary to search the lookup table when obtaining the maximum value in a certain dimension in the lookup table.

(Third embodiment)
A plurality of methods can be applied to the processing performed by the identification contribution degree calculation unit 215 and the feature category selection unit 216 described in the first embodiment. For example, in step S404, the maximum value of the lookup table of each dimension included in the feature category may be obtained, and the variance may be used as the contribution. In this case, in step S405, a feature category having a small variance is selected. A small variance means that the feature category is stable, and an error in approximating the lookup table with the correction value by the feature category correction value calculation unit 217 and the identification value correction unit 225 becomes small.

  Also, for example, as learning sample images, a large number of images that are not identification target objects are collected, and a feature vector distribution for each feature category, that is, a background distribution is obtained. At this time, an information amount for the background distribution is obtained for each feature category, and this information amount, or a remainder in the number of dimensions of the feature category, may be used as the identification contribution. In this case, in step S405, a feature category having a large degree of identification contribution is selected as effective. In the calculation of the information amount, for example, both the background distribution and the distribution of the identification target object are obtained for each feature category as a pair of average and variance, and the KL divergence of both is obtained. A large amount of information means that the probability in the identification value calculation unit 224 is large. Also, a large amount of information per dimension means that there is a high probability per dictionary size. Further, these scales and selection methods may be combined.

(Fourth embodiment)
A plurality of methods can be applied to the processing of the feature category correction value calculation unit 217 and the identification value correction unit 225 described in the first embodiment. For example, assume that the prior distribution of the values of each dimension of the acquired feature vector is uniform, and that the identification values of the selected feature categories and the identification values of the unselected feature categories are correlated. May be. In this case, the average and variance of the maximum values of all dimensions in the selected feature category and the average and variance of the maximum values of all dimensions in the unselected feature category are stored in the dictionary as parameters for correcting the identification value. save. The identification value correction unit 225 first obtains the average of the feature vector values for the selected category. Next, for this average, a Z value in the distribution of maximum values of all dimensions of the selected feature category in the dictionary is obtained. From this Z value, a correction value is estimated based on the mean and variance of unselected feature categories in the dictionary.

  For example, when the kernel function in the classifier learning unit 213 is Histogram Intersection and the discrimination value correction unit 225 intends to minimize false positives, the following may be performed. That is, the correction value may not be stored in the dictionary, and the identification value correction unit 225 may do nothing.

(Fifth embodiment)
The processing of the discriminator learning unit 213 described in the first embodiment is not necessarily a method corresponding to the discriminant value calculation unit 224. Using only the features corresponding to the feature category selected by the feature category selection unit 216, the detector may be learned by a learning method corresponding to the identification method of the identification value calculation unit 224, and the result may be stored in the dictionary. . In this case, the amount of calculation in the learning unit 201 increases, but the identification value correction unit 225 and the like are not necessary, and identification adapted to the selected feature category can be expected.

(Other embodiments)
The preferred embodiment of the present invention has been described in detail above, but the present invention is not limited to the specific embodiment, and various modifications and changes can be made within the scope of the gist of the present invention described in the claims. Is possible.

  The present invention can also be realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiments is supplied to a system or apparatus via a network or various storage media, and a computer (or CPU, MPU, or the like) of the system or apparatus reads the program. It is a process to be executed.

DESCRIPTION OF SYMBOLS 201 Learning part 202 Identification part 211 Feature vector acquisition part 212 1st feature vector normalization part 213 Classifier learning part 214 Look-up table creation part 215 Identification contribution degree calculation part 216 Feature category selection part 217 Feature category correction value calculation part 218 Dictionary creation unit 221 Dictionary reading unit 222 Selected category feature vector acquisition unit 223 Second feature vector normalization unit 224 Discrimination value calculation unit 225 Discrimination value correction unit

Claims (7)

First acquisition means for acquiring a feature vector indicating the characteristics of the learning sample image;
A computing means for obtaining an identification contribution for each feature category from the feature vector of the learning sample image obtained by the first obtaining means;
A selection means for selecting a feature category whose identification contribution obtained by the calculation means is a predetermined value or more;
A dictionary creation means for creating a dictionary including information on the feature category selected by the selection means;
Second acquisition means for acquiring a feature vector indicating the feature of the input image;
Whether the object of the input image is similar to the object of the learning sample image based on the feature vector of the input image acquired by the second acquiring unit and the dictionary generated by the dictionary generating unit Determining means for determining whether or not,
The image identification apparatus characterized in that the second acquisition means does not acquire a feature vector corresponding to a feature category not selected by the selection means from the input image. The arithmetic means extracts a support vector value for each dimension from the feature vector of the learning sample image to obtain a weighted sum related to the support vector, and represents a representative value related to the weighted sum for each predetermined section as the dimension. Means for creating a lookup table stored for each
The image identification apparatus according to claim 1, wherein an identification contribution is calculated for each feature category based on the lookup table.   The image identification apparatus according to claim 2, wherein the calculation unit obtains an average value of maximum values of representative values of each dimension stored in the lookup table as an identification contribution degree of the feature category. The learning sample image is an image that does not belong to a class of objects of the input image;
2. The calculation means obtains an information amount of a background distribution indicating a feature vector distribution for each feature category of the learning sample image acquired by the first acquisition means as the identification contribution degree. Or the image identification apparatus of 2.   The dictionary creation unit includes a lookup table corresponding to a feature category selected by the selection unit among the lookup tables created by the calculation unit, in the dictionary. The image identification device according to 2 or 3. A first acquisition step of acquiring a feature vector indicating the feature of the learning sample image;
A calculation step of obtaining an identification contribution for each feature category from the feature vector of the learning sample image acquired in the first acquisition step;
A selection step of selecting a feature category whose identification contribution determined in the calculation step is a predetermined value or more;
A dictionary creation step of creating a dictionary including information on the feature category selected in the selection step;
A second acquisition step of acquiring a feature vector indicating the feature of the input image;
Whether the object of the input image is similar to the object of the learning sample image based on the feature vector of the input image acquired in the second acquisition step and the dictionary created in the dictionary creation step A determination step for determining whether or not,
In the second acquisition step, a feature vector corresponding to a feature category not selected in the selection step is not acquired from the input image. A first acquisition step of acquiring a feature vector indicating the feature of the learning sample image;
A calculation step of obtaining an identification contribution for each feature category from the feature vector of the learning sample image acquired in the first acquisition step;
A selection step of selecting a feature category whose identification contribution determined in the calculation step is a predetermined value or more;
A dictionary creation step of creating a dictionary including information on the feature category selected in the selection step;
A second acquisition step of acquiring a feature vector indicating the feature of the input image;
Whether the object of the input image is similar to the object of the learning sample image based on the feature vector of the input image acquired in the second acquisition step and the dictionary created in the dictionary creation step A determination step for determining whether or not the computer is executed,
In the second acquisition step, a feature vector corresponding to a feature category not selected in the selection step is not acquired from the input image.

Images