JP5772442B2 - Image processing apparatus and image processing program - Google Patents

Description

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

There is a technology for recognizing images.
As a technology related to this, for example, Patent Document 1 has a plurality of layers for performing parallel bound feature detection for local feature extraction for the purpose of performing highly accurate and reliable optical character recognition. And realized by a hierarchical network having a plurality of fully connected layers for reducing dimensionality, character classification is also performed in the fully connected layer, and each layer of parallel bound feature detection includes a plurality of bound feature maps and corresponding features. It is disclosed that a predetermined kernel is directly associated with a single constrained feature map, and undersampling between layers is performed.

  Also, for example, Patent Document 2 accurately recognizes each character or number of various forms and sizes in order to make a useful decision regarding the identification of characters and numbers using neural network technology. The accuracy in recognizing many letters and numbers is learned so that the neural network can identify the properties belonging to the "invariant" category of each letter or number. It is understood that there is almost no sacrifice, so instead of requiring the neural network to recognize all the stages of invariance identified, such as form, position, size, etc. Requires a much smaller number of sample data entries and little processing of information related to unknown letters or numbers So that a, generalized and restricted described according to the invariant segment that has been disclosed for use.

  Moreover, for example, Patent Document 3 provides a pattern recognition method capable of recognizing not only a two-dimensional object but also a three-dimensional object whose three-dimensional rotation, size, and illumination conditions change. The elements such as weight distribution and pooling stage are the same as the previous method, but it provides a technology that focuses on a new method for determining the optimal feature detection unit at the intermediate stage of the hierarchical network, and also introduces a new feature detection stage (incremental). A new approach to training hierarchical networks using statistical means that learns and significantly reduces the effort required to recognize complex patterns compared to the prior art, and this learning is unsupervised, No teacher signal is required, the recognition architecture can be pre-configured for specific recognition scenarios, and training with supervised learning is required Is only the final classification step, it is disclosed that a considerable effort is reduced in application to thereby recognition task.

JP 05-006463 A Japanese Patent Laid-Open No. 07-064941 JP 2002-373333 A

  The present invention provides an image processing apparatus and an image processing program that reduce identification errors caused by normalization processing when identifying an image, as compared to the case where the present configuration is not provided. The purpose is to do.

The gist of the present invention for achieving the object lies in the inventions of the following items.
According to the first aspect of the present invention, an image receiving means for receiving an image, a plurality of convolution processing means for performing a convolution process on the image received by the image receiving means, and a processing result by the convolution processing means. On the other hand, a plurality of rectification processing means for performing rectification processing, a plurality of normalization processing means for performing normalization processing on a processing result by the rectification processing means, and a processing result by the normalization processing means, A plurality of feature extraction means for extracting feature quantities of the image by performing sub-sampling processing; and an identification means for identifying the image based on the feature quantities extracted by the feature extraction means. The image processing means normalizes the difference between the processing results of the two rectification processing means by dividing the difference based on the value based on the processing results as a divisor. It is a device.

  The invention of claim 2 includes an image receiving means for receiving an image, a plurality of convolution processing means for performing a convolution process on the image received by the image receiving means, and a processing result by the convolution processing means. On the other hand, a plurality of rectification processing means for performing rectification processing, a plurality of sub-sampling processing means for performing sub-sampling processing on a processing result by the rectification processing means, and a processing result by the sub-sampling processing means, A plurality of feature extraction means for extracting feature amounts of the image by performing normalization; and an identification means for identifying the image based on the feature amounts extracted by the feature extraction means. The processing unit divides the difference between the processing results by the two sub-sampling processing units by a value based on the processing results by the two rectification processing units. An image processing apparatus characterized by normalized by dividing by.

  According to a third aspect of the present invention, an image receiving means for receiving an image, a plurality of convolution processing means for performing a convolution process on the image received by the image receiving means, and a processing result by the convolution processing means. On the other hand, a plurality of rectification processing means for performing rectification processing, a plurality of sub-sampling processing means for performing sub-sampling processing on a processing result by the rectification processing means, and a processing result by the sub-sampling processing means, A plurality of feature extraction means for extracting feature amounts of the image by performing normalization; and an identification means for identifying the image based on the feature amounts extracted by the feature extraction means. The processing unit divides the difference between the processing results of the two sub-sampling processing units by using a value based on the processing result as a divisor. An image processing apparatus which is characterized in that-normalized.

  According to a fourth aspect of the present invention, there is provided an image receiving means for receiving an image, a plurality of convolution processing means for performing a convolution process on the image received by the image receiving means, and a processing result by the convolution processing means. On the other hand, a plurality of rectification processing means for performing rectification processing, a plurality of sub-sampling processing means for performing sub-sampling processing on a processing result by the rectification processing means, and a processing result by the sub-sampling processing means, A plurality of feature extraction means for extracting feature amounts of the image by performing normalization; and an identification means for identifying the image based on the feature amounts extracted by the feature extraction means. The normalization processing means normalizes the difference between the processing results of the two sub-sampling processing means by dividing the difference between the sums of the processing results as a divisor. It is an image processing apparatus according to claim.

  The invention of claim 5 includes an image receiving means for receiving an image, a plurality of convolution processing means for performing a convolution process on the image received by the image receiving means, and a processing result by the convolution processing means. On the other hand, a plurality of rectification processing means for performing rectification processing, a plurality of addition means for adding the processing results by the plurality of rectification processing means, and a plurality of sub-sampling processes for processing results by the addition means Sampling processing means, a plurality of averaging processing means for performing averaging processing on the processing results by the sub-sampling processing means, and normalization processing on the processing results by the averaging processing means, A plurality of feature extraction means for extracting the feature quantity of the image; and an identification means for identifying the image based on the feature quantity extracted by the feature extraction means. The processing by the rectification processing means, the adding means, the sub-sampling processing means, and the averaging processing means is a weighted generalized averaging process, and the normalization processing means includes two of the averaging processing means The image processing apparatus is characterized by normalizing the difference between the processing results obtained by subtracting the sum of the processing results as a divisor.

  According to a sixth aspect of the present invention, the rectification processing performed by the rectification processing means is a process (r is a positive real number) for performing an r-th power on an input absolute value, and the averaging processing performed by the averaging processing means is The image processing apparatus according to claim 5, wherein the image processing apparatus performs r ′ power.

  According to a seventh aspect of the present invention, there is provided an image receiving means for receiving an image, a plurality of convolution processing means for performing a convolution process on the image received by the image receiving means, and the convolution processing means. A plurality of rectification processing means for performing rectification processing on the processing result, a plurality of normalization processing means for performing normalization processing on a processing result by the rectification processing means, and a processing result by the normalization processing means. On the other hand, by performing sub-sampling processing, a plurality of feature extraction means for extracting the feature quantity of the image, and an identification means for identifying the image based on the feature quantity extracted by the feature extraction means. The normalization processing means normalizes the difference between the processing results of the two rectification processing means by dividing the value based on the processing results as a divisor. An image processing program characterized and.

  According to an eighth aspect of the present invention, the computer includes an image receiving unit that receives an image, a plurality of convolution processing units that perform a convolution process on the image received by the image receiving unit, and the convolution processing unit. A plurality of rectification processing means for performing rectification processing on the processing result, a plurality of sub-sampling processing means for performing sub-sampling processing on a processing result by the rectification processing means, and a processing result by the sub-sampling processing means On the other hand, by performing normalization processing, a plurality of feature extraction means for extracting the feature amount of the image, and an identification means for identifying the image based on the feature amount extracted by the feature extraction means. The normalization processing means calculates the difference between the processing results of the two sub-sampling processing means by the two rectification processing means. An image processing program characterized by normalizing by dividing a value based on results as divisor.

  According to a ninth aspect of the present invention, there is provided a computer comprising: an image receiving unit that receives an image; a plurality of convolution processing units that perform a convolution process on the image received by the image receiving unit; and the convolution processing unit. A plurality of rectification processing means for performing rectification processing on the processing result, a plurality of sub-sampling processing means for performing sub-sampling processing on a processing result by the rectification processing means, and a processing result by the sub-sampling processing means On the other hand, by performing normalization processing, a plurality of feature extraction means for extracting the feature amount of the image, and an identification means for identifying the image based on the feature amount extracted by the feature extraction means. The normalization processing means uses the difference between the processing results of the two sub-sampling processing means as a divisor with a value based on the processing results. An image processing program characterized by normalized by dividing.

  According to a tenth aspect of the present invention, the computer includes an image receiving unit that receives an image, a plurality of convolution processing units that perform a convolution process on the image received by the image receiving unit, and the convolution processing unit. A plurality of rectification processing means for performing rectification processing on the processing result, a plurality of sub-sampling processing means for performing sub-sampling processing on a processing result by the rectification processing means, and a processing result by the sub-sampling processing means On the other hand, by performing normalization processing, a plurality of feature extraction means for extracting the feature amount of the image, and an identification means for identifying the image based on the feature amount extracted by the feature extraction means. The normalization processing means divides the difference between the processing results of the two sub-sampling processing means by using the sum of the processing results as a divisor. An image processing program characterized by normalized by Rukoto.

  According to an eleventh aspect of the present invention, the computer includes an image receiving unit that receives an image, a plurality of convolution processing units that perform a convolution process on the image received by the image receiving unit, and the convolution processing unit. A plurality of rectification processing means for performing a rectification process on the processing result, a plurality of addition means for adding the processing results by the plurality of rectification processing means, and a sub-sampling process on the processing result by the addition means A plurality of sub-sampling processing means, a plurality of averaging processing means for performing an averaging process on the processing result by the sub-sampling processing means, and a normalization process for the processing result by the averaging processing means A plurality of feature extraction means for extracting the feature quantity of the image, and the image based on the feature quantity extracted by the feature extraction means. The processing by the rectification processing unit, the addition unit, the sub-sampling processing unit, and the averaging processing unit is a weighted generalized averaging process, and the normalization processing unit includes two normalization processing units. An image processing program characterized by normalizing a difference between processing results obtained by the averaging processing means by dividing a sum of the processing results as a divisor.

  According to the image processing apparatus of the first aspect, when identifying an image, it is possible to reduce identification errors caused by normalization as compared with a case where the present configuration is not provided.

  According to the image processing apparatus of the second aspect, when the image is identified, the amount of calculation of the subtraction process and the division process in the normalization process can be reduced as compared with the case where the present configuration is not provided. it can.

  According to the image processing apparatus of the third aspect, it is possible to reduce the amount of calculation in the normalization process when identifying an image as compared with the case where the present configuration is not provided.

  According to the image processing apparatus of the fourth aspect, it is possible to reduce the amount of calculation in the normalization process when an image is identified and compared with a case where this configuration is not provided.

  According to the image processing apparatus of the fifth aspect, when the image is identified, the learning time can be shortened as compared with the case where the present configuration is not provided.

  According to the image processing apparatus of the sixth aspect, when the image is identified, the learning time can be shortened as compared with the case where the present configuration is not provided.

  According to the image processing program of the seventh aspect, it is possible to reduce the identification error caused by the normalization process when the image is identified and compared with the case where the present configuration is not provided.

  According to the image processing program of the eighth aspect, when the image is identified, the calculation amount of the subtraction process and the division process in the normalization process can be reduced as compared with the case where the present configuration is not provided. it can.

  According to the image processing program of the ninth aspect, it is possible to reduce the amount of calculation in the normalization process in comparison with the case where the image is identified and this configuration is not provided.

  According to the image processing program of the tenth aspect, it is possible to reduce the amount of calculation in the normalization process when the image is identified and compared with the case where the present configuration is not provided.

  According to the image processing program of the eleventh aspect, it is possible to reduce the learning time when identifying an image as compared with the case where the present configuration is not provided.

It is a conceptual module block diagram about the structural example of 1st Embodiment. It is a conceptual module block diagram about the example of a whole structure of 1st Embodiment. It is a conceptual module block diagram about the structural example of 2nd Embodiment. It is a conceptual module block diagram about the structural example of 3rd Embodiment. It is a notional module block diagram about the structural example of 4th Embodiment. It is a notional module block diagram about the structural example of 5th Embodiment. It is a notional module block diagram about the structural example of 6th Embodiment. It is a block diagram which shows the hardware structural example of the computer which implement | achieves this Embodiment. It is a conceptual module block diagram about the example of a whole structure of a CNN identification device. It is a conceptual module block diagram about the example of the feature extraction layer which performs a rectification | straightening process and a normalization process. It is explanatory drawing which shows the example of a normalization process. It is explanatory drawing which shows the example of a normalization process. It is explanatory drawing which shows the example of a normalization process.

First, before describing the present embodiment, the premise or an image processing apparatus using the present embodiment will be described. This description is intended to facilitate understanding of the present embodiment.
There is a technique for identifying an image into one of a plurality of classes. This is so-called recognition technology. Also, among these, there is one that learns an identification device from a set of an image and a correct class and improves the recognition rate. Specifically, this is a technique related to a hierarchical network for identifying an accepted pattern (image).
In particular, it is known that an image can be identified using a convolutional neural network (hereinafter referred to as CNN) which is a kind of hierarchical neural network. [LeCun, Yann, et al. “Gradient-Based Learning Applied to Document Recognition.” Proceedings of the IEEE 86.11 (1998): 2278-2324. ]

FIG. 9 is a conceptual module configuration diagram of an overall configuration example of the CNN identification device.
The feature extraction layer (feature extraction layer 910, feature extraction layer 920) of the CNN identification device is composed of a convolution layer and a sub-sampling layer. The input of each layer is a three-dimensional feature map composed of feature quantities arranged in a plurality of two-dimensional spaces, and the output is a new feature map. The image 900 received by the CNN identification device is also regarded as a kind of feature map.
In the convolution layer, weights are convolved with the received feature amount (for example, a two-dimensional digital filter is used as the convolution process), and in the sub-sampling layer, the response of the convolution is sub-sampled. The obtained feature quantity is converted into a feature quantity that is invariant to local variations in position.
This feature extraction layer (feature extraction layer 910, feature extraction layer 920) is repeated a plurality of times (in FIG. 9, although it is two feature extraction layers, it may be three or more), and the output of the classifier 930 is Accept.
The discriminator 930 identifies the class to which the image 900 belongs from the received feature amount. The discriminator 930 is generally composed of a multilayer perceptron or an RBF (Radial Basis Function).
The convolution weights of the feature extraction layers (the feature extraction layer 910 and the feature extraction layer 920) are generally performed by an error back propagation method using a set of images and correct answer classes.
For example, JP-A-5-6463, JP-A-7-64941, JP-A-2002-373333, JP-A-2005-215988, JP-A-2010-157118, and the like can be cited as references.

  In recent years, in order to further improve the identification accuracy, rectification processing (Rectification) and normalization processing (Division Normalization / Contrast Normalization) have been added to the feature extraction layer of CNN [LeCun, Yann, Korykkubumulu. “Convolutional Networks and Applications in Vision.” Proceedings of 2010 IEEE International Symposium on Circuits and Systems (2010): 253-256. ]. Hereinafter, the term “CNN” refers to a device to which this rectification processing and normalization processing are added.

FIG. 10 is a conceptual module configuration diagram of an example of a feature extraction layer that performs rectification processing and normalization processing. FIG. 10 illustrates the case where there are two convolution processing modules 1010 and the like, but may be three or more.
The feature extraction layer includes a plurality of convolution processing modules 1010, a plurality of rectification processing modules 1020, a normalization processing module 1030, and a plurality of sub-sampling processing modules 1040. That is, rectification processing and normalization processing are added to the conventional CNN.
The convolution processing module 1010a and the convolution processing module 1010b are connected to the rectification processing module 1020a and the rectification processing module 1020b, respectively, and receive images or feature maps 1000A, 1000B, and 1000C and perform convolution processing on these. .
The rectification processing module 1020a is connected to the convolution processing module 1010a, the average calculation processing module 1032, the difference calculation processing module 1034a, and the standard deviation calculation processing module 1036. In the conventional CNN, feature cancellation occurs due to sub-sampling between convolutional responses in which positive and negative are mixed. In order to improve the recognition rate, this rectification process is added to suppress the cancellation of features. Specific examples of the rectification process include an absolute value process and a square process.
The rectification processing module 1020b is connected to the convolution processing module 1010b, the average calculation processing module 1032, the difference calculation processing module 1034b, and the standard deviation calculation processing module 1036, and performs processing equivalent to that of the rectification processing module 1020a.

The normalization processing module 1030 includes an average calculation processing module 1032, a difference calculation processing module 1034a, a difference calculation processing module 1034b, a standard deviation calculation processing module 1036, a division processing module 1038a, and a division processing module 1038b. In order to improve the recognition rate, by adding this normalization process, even if the feature value accepted by convolution is displaced uniformly, it is possible to obtain a stable feature value by suppressing fluctuations. .
The average calculation processing module 1032 is connected to the rectification processing module 1020a, the rectification processing module 1020b, the difference calculation processing module 1034a, and the difference calculation processing module 1034b. The average calculation processing module 1032 calculates an average value of all the values after the rectification processing, and passes the average value to the difference calculation processing module 1034a and the like.
The difference calculation processing module 1034a is connected to the rectification processing module 1020a, the average calculation processing module 1032 and the division processing module 1038a. The difference calculation processing module 1034a performs processing for subtracting the average value calculated by the average calculation processing module 1032 from the processing result of the rectification processing module 1020a.
The difference calculation processing module 1034b is connected to the rectification processing module 1020b, the average calculation processing module 1032 and the division processing module 1038b, and performs processing equivalent to the difference calculation processing module 1034a.
The standard deviation calculation processing module 1036 is connected to the rectification processing module 1020a, the rectification processing module 1020b, the division processing module 1038a, and the division processing module 1038b. The standard deviation calculation processing module 1036 calculates the standard deviation of all the values after the rectification processing, and passes the standard deviation to the division processing module 1038a and the like.
The division processing module 1038a is connected to the difference calculation processing module 1034a, the standard deviation calculation processing module 1036, and the sub-sampling processing module 1040a. The division processing module 1038a divides the processing result by the difference calculation processing module 1034a by the standard deviation calculated by the standard deviation calculation processing module 1036 (that is, the standard deviation is a divisor).
The division processing module 1038b is connected to the difference calculation processing module 1034b, the standard deviation calculation processing module 1036, and the subsampling processing module 1040b, and performs processing equivalent to the division processing module 1038a.
The sub-sampling processing module 1040a is connected to the division processing module 1038a. The sub-sampling processing module 1040a performs sub-sampling processing on the processing result obtained by the division processing module 1038a. Then, it is output to the next feature extraction layer (that is, an image or feature map 1000A, 1000B, 1000C, etc. for the next feature extraction layer) or output to a discriminator (also referred to as a discrimination layer).
The sub-sampling processing module 1040b is connected to the division processing module 1038b and performs processing equivalent to the sub-sampling processing module 1040a.

In the normalization processing performed by the normalization processing module 1030, the received feature amount is scaled and output so as to take a value within a certain range (for example, [−1, 1]). FIG. 11 is an explanatory diagram illustrating an example of normalization processing. The graph shown in the example of FIG. 11A is a value accepted by the normalization processing module 1030, and the graph shown in the example of FIG. 11B is a value output by the normalization processing module 1030. In this case, the enlargement processing It is carried out.
For example, when the variance of the values received by the normalization processing module 1030 is a small value, the value is greatly expanded by the normalization processing, so the output value is the sign of the normalization denominator, that is, the average value of the received values. Strongly dependent.
Further, when a part of the values received by the normalization processing module 1030 takes a large value (for example, it may be a large value due to the influence of noise or the like), the entire output value changes greatly.
12 and 13 are explanatory diagrams showing an example of normalization processing in such a case. The graph shown in the example of FIG. 12A is a value accepted by the normalization processing module 1030, and a large value is initially generated due to the influence of noise or the like. Therefore, the average value of these values is higher than the average value calculated with other values. The graph shown in the example of FIG. 12B is a value output from the normalization processing module 1030 in this case.
The graph shown in the example of FIG. 13A is a value accepted by the normalization processing module 1030. Compared with the graph shown in the example of FIG. 12A, a large value is not generated due to the influence of noise or the like, but the graph is similar except for the first part of the example of FIG. The average value of this value is lower than the average value in the example of FIG. The graph shown in the example of FIG. 13B is a value output from the normalization processing module 1030 in this case. That is, although the graph shown in the example of FIG. 12B and the graph shown in the example of FIG. 13A are similar, the graph shown in the example of FIG. 12B and the graph of FIG. The graphs shown in the examples are different. That is, the output of the normalization process depends on the average value of the accepted values.
As described above, in the feature extraction layer of the CNN, a variation in a part of values (a part of the convolution filter response) accepted by the normalization process may cause a large variation in the entire output feature amount. That is, there may be a case where a stable feature amount cannot be obtained with respect to fluctuations in the received image.

Hereinafter, examples of various preferred embodiments for realizing the present invention will be described with reference to the drawings.
<First Embodiment>
FIG. 1 is a conceptual module configuration diagram of a configuration example according to the first embodiment.
The module generally refers to components such as software (computer program) and hardware that can be logically separated. Therefore, the module in the present embodiment indicates not only a module in a computer program but also a module in a hardware configuration. Therefore, the present embodiment is a computer program for causing these modules to function (a program for causing a computer to execute each procedure, a program for causing a computer to function as each means, and a function for each computer. This also serves as an explanation of the program and system and method for realizing the above. However, for the sake of explanation, the words “store”, “store”, and equivalents thereof are used. However, when the embodiment is a computer program, these words are stored in a storage device or stored in memory. It is the control to be stored in the device. Modules may correspond to functions one-to-one, but in mounting, one module may be configured by one program, or a plurality of modules may be configured by one program, and conversely, one module May be composed of a plurality of programs. The plurality of modules may be executed by one computer, or one module may be executed by a plurality of computers in a distributed or parallel environment. Note that one module may include other modules. Hereinafter, “connection” is used not only for physical connection but also for logical connection (data exchange, instruction, reference relationship between data, etc.). “Predetermined” means that the process is determined before the target process, and not only before the process according to this embodiment starts but also after the process according to this embodiment starts. In addition, if it is before the target processing, it is used in accordance with the situation / state at that time or with the intention to be decided according to the situation / state up to that point. In addition, the description having the meaning of “do B when it is A” is used in the meaning of “determine whether or not it is A and do B when it is judged as A”. However, the case where it is not necessary to determine whether or not A is excluded.
In addition, the system or device is configured by connecting a plurality of computers, hardware, devices, and the like by communication means such as a network (including one-to-one correspondence communication connection), etc., and one computer, hardware, device. The case where it implement | achieves by etc. is included. “Apparatus” and “system” are used as synonymous terms. Of course, the “system” does not include a social “mechanism” (social system) that is an artificial arrangement.
In addition, when performing a plurality of processes in each module or in each module, the target information is read from the storage device for each process, and the processing result is written to the storage device after performing the processing. is there. Therefore, description of reading from the storage device before processing and writing to the storage device after processing may be omitted. Here, the storage device may include a hard disk, a RAM (Random Access Memory), an external storage medium, a storage device via a communication line, a register in a CPU (Central Processing Unit), and the like.

The image processing apparatus according to the first embodiment extracts features for identifying an image. As shown in the example of FIG. 1, the convolution processing module 110, the rectification processing module 120, and the normalization are performed. A processing module 130 and a sub-sampling processing module 140 are included.
This image processing apparatus is a feature extraction layer, and for a pair of feature values obtained by performing rectification processing on responses of two convolution filters, a difference in the pair is calculated by dividing a value based on the feature value in the pair. As a result of normalization by dividing as a sub-sampled output feature value.

The convolution processing module 110a, the convolution processing module 110b, the convolution processing module 110c, and the convolution processing module 110d are connected to the rectification processing module 120a, the rectification processing module 120b, the rectification processing module 120c, and the rectification processing module 120d, respectively. . The convolution processing module 110 receives images or feature maps 100A, 100B, and 100C. In the case of the feature extraction layer in the first stage, an image is accepted, and in the case of the feature extraction layers in the second and subsequent stages, a feature amount that is an output in the preceding feature extraction layer is accepted. Then, a convolution process is performed on the received image or feature amount. Processing equivalent to that of the convolution processing module 1010 illustrated in FIG. 10 described above is performed.
The rectification processing module 120a is connected to the convolution processing module 110a, the difference calculation processing module 132a, and the divisor calculation processing module 134a. The rectification processing module 120a performs rectification processing. Processing equivalent to that of the rectification processing module 1020 illustrated in FIG. 10 described above is performed. Of course, the processing of the rectification processing module 1020b and the like is the same.

The normalization processing module 130 includes a difference calculation processing module 132a, a difference calculation processing module 132c, a divisor calculation processing module 134a, a divisor calculation processing module 134c, a division processing module 136a, and a division processing module 136c. The normalization processing module 130 performs normalization processing.
The difference calculation processing module 132a is connected to the rectification processing module 120a, the rectification processing module 120b, and the division processing module 136a.
The divisor calculation processing module 134a is connected to the rectification processing module 120a, the rectification processing module 120b, and the division processing module 136a.
The division processing module 136a is connected to the difference calculation processing module 132a, the divisor calculation processing module 134a, and the subsampling processing module 140a.
As a normalization process performed by the normalization processing module 130, a difference between the processing results of the two rectification processing modules 120 (for example, a pair of the rectification processing module 120a and the rectification processing module 120b) is determined based on the processing result (for example, , Average value, standard deviation, mode value, median value, etc.) are normalized as a divisor. Note that the difference between the processing results is that the other processing result is subtracted from one processing result, and either the rectification processing module 120a or the rectification processing module 120b is one, and the other may be the other.
In this feature extraction layer, the influence on the output feature amount caused by the variation in the processing result (filter response) by some convolution processing modules 110 is limited within the pair, and some filter responses are output. It suppresses large fluctuations in the entire feature value.
Further, by taking the difference between the processing results of the pair, the CNN can obtain a stable feature amount with suppressed fluctuation even when the feature amount received by the convolution processing module 110 is displaced large (or small). The advantage of is kept.
In particular, when the value of the image or feature map is 0, the feature amount output by the feature extraction layer is also 0, and accordingly, it is confirmed that the received image or feature map value is 0. Will be communicated.

  The sub-sampling processing module 140a is connected to the division processing module 136a. By performing a sub-sampling process on the processing result by the normalization processing module 130, the feature amounts of the images or feature maps 100A, 100B, and 100C are extracted. And it outputs to the feature extraction layer of the next stage. Or it outputs to a discriminator. The discriminator identifies the images or feature maps 100A, 100B, and 100C based on the received feature amount.

  FIG. 2 is a conceptual module configuration diagram of an overall configuration example of the first embodiment. This is a stack of the feature extraction layers illustrated in FIG. 1 in three layers. As the feature extraction layer 1: 210, the feature extraction layer 2: 230, and the feature extraction layer 3: 250, respectively, the outputs are the feature map A. : 220, feature map B: 240, and feature map C: 260. Of course, it is not limited to three layers, but may be two layers or three or more layers. Note that the feature extraction layer (image processing apparatus) of the second to sixth embodiments can also be applied to the feature extraction layer illustrated in FIG.

となる。なお、ｉは出力特徴量マップのインデックス、ｐは入力特徴量マップのインデックス、ｊ，ｋは特徴量マップ内の２次元座標、ｑ、ｒは畳込みフィルタの２次元座標である。 When the feature quantity received by the feature extraction layer illustrated in FIG. 1 is x and the convolution filter is w, the output feature quantity s of the convolution process is

It becomes. Note that i is an index of the output feature map, p is an index of the input feature map, j and k are two-dimensional coordinates in the feature map, and q and r are two-dimensional coordinates of the convolution filter.

の関係がある。 When the size of the input feature map for convolution processing is M × N, the size of the output feature map is M ′ × N ′, and the size of the convolution filter is Q × R,

There is a relationship.

となる。なお、ｉは特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標である。 In the feature extraction layer illustrated in FIG. 1, when the output feature quantity of the convolution process is s and the output feature quantity of the rectification process is t,

It becomes. Note that i is a feature map index, and j and k are two-dimensional coordinates in the feature map.

となる。なお、ｉは出力特徴量マップのインデックス、ｐは入力特徴領マップのインデックス、ｊ，ｋは特徴量マップ内の２次元座標、ｑ、ｒは除数を計算する局所領域のインデックス、ｗ及びβは予め与えられる重み付き一般化平均のパラメータである。 In the feature extraction layer illustrated in FIG. 1, when the output feature amount of the rectification process is t and the output feature amount of the divisor calculation process is v,

It becomes. Note that i is an index of the output feature map, p is an index of the input feature map, j and k are two-dimensional coordinates in the feature map, q and r are indices of a local region for calculating a divisor, and w and β are This is a weighted generalized average parameter given in advance.

となる。なお、ｉは出力特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標である。 In the feature extraction layer illustrated in FIG. 1, when the output feature amount of the rectification process is t and the output feature amount of the normalization process is u,

It becomes. Note that i is an index of the output feature map, and j and k are two-dimensional coordinates in the feature map.

となる。なお、ｉは出力特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標、ｑ、ｒは除数を計算する局所領域のインデックス、ｗは予め与えられるサブサンプリングの重みである。 In the feature extraction layer illustrated in FIG. 1, assuming that the output feature amount of the normalization process is u and the output of the feature extraction layer is y,

It becomes. Note that i is an index of the output feature map, j and k are two-dimensional coordinates in the feature map, q and r are indices of local regions for calculating the divisor, and w is a weight of subsampling given in advance.

の計算で構成される。なお、ｗは学習されるウェイトである。出力の添え字ｃは入力画像が識別されるクラスを表し、識別時にはｚ_ｃがもっとも大きな値となるｃが識別結果となる。
ウェイトの学習は、受け付ける画像と正解識別クラスの組を用いて、誤差逆伝搬法にてクロスエントロピー誤差を最小化させて行われる。
したがって、識別器は、１層の全結合層とＳｏｆｔｍａｘ関数によって構成されていてもよい。識別器は、１層の全結合層とＲＢＦ関数によって構成されていてもよい。特徴抽出層及び識別器は、画像と正解クラスの組から結合重みを学習するようにしてもよい。結合重みの学習は、誤差逆伝搬法にて行われるようにしてもよい。 The classifier that obtains the final output z from the output x of the feature map C: 260 illustrated in FIG.

It consists of the calculation of Note that w is a weight to be learned. The subscript c of the output represents the class in which the input image is identified, and at the time of identification, _c having the largest value of zc is the identification result.
Weight learning is performed by minimizing a cross-entropy error by a back propagation method using a set of an image to be received and a correct answer identification class.
Therefore, the discriminator may be composed of one full coupling layer and a Softmax function. The discriminator may be composed of one full coupling layer and an RBF function. The feature extraction layer and the discriminator may learn the connection weight from the set of the image and the correct class. The learning of the connection weight may be performed by an error back propagation method.

<Second Embodiment>
FIG. 3 is a conceptual module configuration diagram of a configuration example according to the second embodiment. The processing result according to the first embodiment is not changed, and in order to reduce the calculation amount of subtraction and division of normalization processing, the subsampling processing of the feature extraction layer illustrated in FIG. Move after normalization processing but before normalization processing.
As shown in the example of FIG. 3, the second embodiment includes a convolution processing module 310, a rectification processing module 320, a sub-sampling processing module 330, and a normalization processing module 340.
The convolution processing module 310 performs processing equivalent to that of the convolution processing module 110.
The rectification processing module 320 is connected to the convolution processing module 310, the sub-sampling processing module 330, and the divisor calculation processing module 344, and performs processing equivalent to that of the rectification processing module 120. To the divisor calculation processing module 344.
The sub-sampling processing module 330 is connected to the rectification processing module 320 and the difference calculation processing module 342. The sub-sampling processing module 330 performs sub-sampling processing on the processing result from the rectification processing module 320.

The normalization processing module 340 includes a difference calculation processing module 342a, a difference calculation processing module 342c, a divisor calculation processing module 344a, a divisor calculation processing module 344c, a division processing module 346a, and a division processing module 346c. The normalization processing module 340 performs the normalization processing on the processing results obtained by the plurality of sub-sampling processing modules 330, thereby extracting the feature amounts of the images or feature maps 100A, 100B, and 100C. And it outputs to the feature extraction layer of the next stage. Or it outputs to a discriminator. The discriminator identifies the images or feature maps 100A, 100B, and 100C based on the received feature amount.
The difference calculation processing module 342a is connected to the sub-sampling processing module 330a, the sub-sampling processing module 330b, and the division processing module 346a.
The divisor calculation processing module 344a is connected to the rectification processing module 320a, the rectification processing module 320b, and the division processing module 346a.
The division processing module 346a is connected to the difference calculation processing module 342a and the divisor calculation processing module 344a.
As normalization processing performed by the normalization processing module 340, the difference between the processing results of the two sub-sampling processing modules 330 (for example, a pair of the sub-sampling processing module 330a and the sub-sampling processing module 330b) is calculated as the two rectification processing modules 320. Normalization is performed by dividing a value (for example, an average value, a standard deviation, a mode value, a median value, and the like) based on the processing result by the above as a divisor. Note that the difference between the processing results is that one processing result is subtracted from the other processing result, and either the sub-sampling processing module 330a or the sub-sampling processing module 330b is one and may be the other. .

なお、ｔは整流処理の出力特徴量、ｕはサブサンプリング処理の出力、ｖは正規化処理の除数、ｙは特徴抽出層の出力、ｉ、ｐは特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標、ｑ、ｒは畳込みフィルタの２次元座標、ｗ及びβは予め与えられる重み付き一般化平均のパラメータである。 In this case, (Expression 4) to (Expression 6) can be rewritten as (Expression 8).

T is the output feature value of the rectification process, u is the output of the sub-sampling process, v is the divisor of the normalization process, y is the output of the feature extraction layer, i and p are the indexes of the feature map, and j and k are the features. Two-dimensional coordinates in the quantity map, q and r are two-dimensional coordinates of the convolution filter, and w and β are parameters of a weighted generalized average given in advance.

<Third Embodiment>
FIG. 4 is a conceptual module configuration diagram of a configuration example according to the third embodiment. In the second embodiment, the input to the divisor calculation process is changed from the feature quantity before sub-sampling to the feature quantity after sub-sampling in order to reduce the amount of calculation of the divisor calculation process of the normalization process. .
As shown in the example of FIG. 4, the third embodiment includes a convolution processing module 410, a rectification processing module 420, a sub-sampling processing module 430, and a normalization processing module 440.
The convolution processing module 410 performs processing equivalent to that of the convolution processing module 310.
The rectification processing module 420 is connected to the convolution processing module 410 and the subsampling processing module 430. Processing equivalent to that of the rectification processing module 320 is performed.
The sub-sampling processing module 430 is connected to the rectification processing module 420, the difference calculation processing module 442, and the divisor calculation processing module 444. The same processing as the sub-sampling processing module 330 is performed, but the processing result is passed to the difference calculation processing module 442 and the divisor calculation processing module 444.

The normalization processing module 440 includes a difference calculation processing module 442a, a difference calculation processing module 442c, a divisor calculation processing module 444a, a divisor calculation processing module 444c, a division processing module 446a, and a division processing module 446c. The normalization processing module 440 extracts the feature amounts of the images or feature maps 100A, 100B, and 100C by performing normalization processing on the processing results of the plurality of sub-sampling processing modules 430. And it outputs to the feature extraction layer of the next stage. Or it outputs to a discriminator. The discriminator identifies the images or feature maps 100A, 100B, and 100C based on the received feature amount.
The difference calculation processing module 442a is connected to the sub-sampling processing module 430a, the sub-sampling processing module 430b, and the division processing module 446a.
The divisor calculation processing module 444a is connected to the sub-sampling processing module 430a, the sub-sampling processing module 430b, and the division processing module 446a.
The division processing module 446a is connected to the difference calculation processing module 442a and the divisor calculation processing module 444a.
As the normalization processing performed by the normalization processing module 440, the difference between the processing results of the two sub-sampling processing modules 430 (for example, a pair of the sub-sampling processing module 430a and the sub-sampling processing module 430b) is based on the processing results. Normalize by dividing the value (eg, mean, standard deviation, mode, median, etc.) as a divisor. Note that the difference between the processing results is that the other processing result is subtracted from one processing result, and either the sub-sampling processing module 430a or the sub-sampling processing module 430b is one and may be the other. .

なお、ｔは整流処理の出力特徴量、ｕはサブサンプリング処理の出力、ｖは正規化処理の除数、ｙは特徴抽出層の出力、ｉ、ｐは特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標、ｑ、ｒは畳込みフィルタの２次元座標、βは予め与えられる重み付き一般化平均のパラメータである。 In this case, (Equation 8) is rewritten as (Equation 9).

T is the output feature value of the rectification process, u is the output of the sub-sampling process, v is the divisor of the normalization process, y is the output of the feature extraction layer, i and p are the indexes of the feature map, and j and k are the features. Two-dimensional coordinates in the quantity map, q and r are two-dimensional coordinates of the convolution filter, and β is a weighted generalized average parameter given in advance.

<Fourth embodiment>
FIG. 5 is a conceptual module configuration diagram of a configuration example according to the fourth embodiment. To reduce the amount of normalization processing, in the third embodiment, when β = 1 in (Equation 9), calculations of the β power and 1 / β power are not required.
As shown in the example of FIG. 5, the fourth embodiment includes a convolution processing module 510, a rectification processing module 520, a sub-sampling processing module 530, and a normalization processing module 540.
The convolution processing module 510 performs processing equivalent to the convolution processing module 410.
The rectification processing module 520 performs processing equivalent to that of the rectification processing module 420.
The subsampling processing module 530 performs processing equivalent to that of the subsampling processing module 430.

The normalization processing module 540 includes a difference calculation processing module 542a, a difference calculation processing module 542c, a sum calculation processing module 544a, a sum calculation processing module 544c, a division processing module 546a, and a division processing module 546c. The normalization processing module 540 extracts the feature amounts of the images or feature maps 100A, 100B, and 100C by performing normalization processing on the processing results of the plurality of sub-sampling processing modules 530. And it outputs to the feature extraction layer of the next stage. Or it outputs to a discriminator. The discriminator identifies the images or feature maps 100A, 100B, and 100C based on the received feature amount.
The difference calculation processing module 542a is connected to the sub-sampling processing module 530a, the sub-sampling processing module 530b, and the division processing module 546a.
The sum calculation processing module 544a is connected to the sub-sampling processing module 530a, the sub-sampling processing module 530b, and the division processing module 546a.
The division processing module 546a is connected to the difference calculation processing module 542a and the sum calculation processing module 544a.
As the normalization processing performed by the normalization processing module 540, the difference between the processing results of the two sub-sampling processing modules 530 (for example, a pair of the sub-sampling processing module 530a and the sub-sampling processing module 530b) Normalize by dividing as a divisor. The difference between the processing results is to subtract the other processing result from one processing result, and either the sub-sampling processing module 530a or the sub-sampling processing module 530b is one and may be the other. .

なお、ｔは整流処理の出力特徴量、ｕはサブサンプリング処理の出力、ｙは特徴抽出層の出力、ｉ、ｐは特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標、ｑ、ｒは畳込みフィルタの２次元座標である。 In this case, (Equation 9) is rewritten as (Equation 10).

T is an output feature value of the rectification process, u is an output of the sub-sampling process, y is an output of the feature extraction layer, i and p are indexes of the feature quantity map, j and k are two-dimensional coordinates in the feature quantity map, q and r are two-dimensional coordinates of the convolution filter.

<Fifth embodiment>
FIG. 6 is a conceptual module configuration diagram of an exemplary configuration according to the fifth embodiment. In the third embodiment and the fourth embodiment, the sub-sampling process can be extended to a generalized average that receives processing results from a plurality of convolution processing modules.
As shown in the example of FIG. 6, the fifth embodiment includes a convolution processing module 610, a rectification + subsampling processing module 620, and a normalization processing module 630.
The convolution processing module 610 performs processing equivalent to the convolution processing module 510.

The rectification + subsampling processing module 620 includes a rectification processing module 622a, a rectification processing module 622b, a rectification processing module 622c, a rectification processing module 622d, a rectification processing module 622e, a rectification processing module 622f, a sum calculation processing module 624a, and a sum calculation processing module 624d. A sub-sampling processing module 626a, a sub-sampling processing module 626d, an averaging processing module 628a, and an averaging processing module 628d.
The rectification processing module 622a is connected to the convolution processing module 610a and the sum calculation processing module 624a. The rectification processing module 622a performs rectification processing on the processing result of the convolution processing module 610a. For example, the absolute value of the processing result by the convolution processing module 610a is calculated to the power of r.
The sum calculation processing module 624a is connected to the rectification processing module 622a, the rectification processing module 622b, the rectification processing module 622c, and the sub-sampling processing module 626a. The sum calculation processing module 624a adds the processing results of the plurality of rectification processing modules 622 (r rectification processing modules 622a, 622b,..., 622c).
The sub-sampling processing module 626a is connected to the sum calculation processing module 624a and the averaging processing module 628a. The sub-sampling processing module 626a performs sub-sampling processing on the processing result by the sum calculation processing module 624a.
The averaging processing module 628a is connected to the sub-sampling processing module 626a, the difference calculation processing module 632, and the sum calculation processing module 634. The averaging processing module 628a performs an averaging process on the processing result obtained by the sub-sampling processing module 626a. For example, the r-th root of the processing result by the sub-sampling processing module 626a is calculated.

The normalization processing module 630 includes a difference calculation processing module 632, a sum calculation processing module 634, and a division processing module 636. The normalization processing module 630 extracts the feature amounts of the images or feature maps 100A, 100B, and 100C by performing normalization processing on the processing results of the plurality of averaging processing modules 628. And it outputs to the feature extraction layer of the next stage. Or it outputs to a discriminator. The discriminator identifies the images or feature maps 100A, 100B, and 100C based on the received feature amount.
The difference calculation processing module 632 is connected to the averaging processing module 628a, the averaging processing module 628d, and the division processing module 636.
The sum calculation processing module 634 is connected to the averaging processing module 628a, the averaging processing module 628d, and the division processing module 636.
The division processing module 636 is connected to the difference calculation processing module 632 and the sum calculation processing module 634.
As the normalization processing performed by the normalization processing module 630, the difference between the processing results of the two averaging processing modules 628 (for example, the averaging processing module 628a and the averaging processing module 628d) is used, and the sum of the processing results is used as a divisor. Normalize by dividing. Note that the difference between the processing results is that one processing result is subtracted from the other processing result, and either the averaging processing module 628a or the averaging processing module 628d is one and may be the other. .

なお、ｓは畳込み処理の出力特徴量、ｕは整流及びサブサンプリング処理の出力、ｙは特徴抽出層の出力、ｉ、ｐは特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標、ｑ、ｒは畳込みフィルタの２次元座標、ｗ及びβは予め与えられる重み付き一般化平均のパラメータである。
正規化処理は、入力の重み付き一般化平均を除数としてもよい（（式１１）の第１式）。 In this case, (Expression 3) and (Expression 10) are rewritten as (Expression 11).

Note that s is an output feature value of the convolution process, u is an output of the rectification and sub-sampling process, y is an output of the feature extraction layer, i and p are feature map indices, and j and k are 2 in the feature map. Dimensional coordinates, q and r are two-dimensional coordinates of the convolution filter, and w and β are parameters of a weighted generalized average given in advance.
In the normalization process, a weighted generalized average of the input may be used as a divisor (the first expression of (Expression 11)).

<Sixth Embodiment>
FIG. 7 is a conceptual module configuration diagram of a configuration example according to the sixth embodiment. In the fifth embodiment, the number of convolution filter responses input to the sub-sampling process is set to 2 and r = 2 in order to obtain an output that is invariant to local phase fluctuations of the image. (Adelson, Edward H and James R Bergen. “Spatiotemporal energy models for the perception of motion. 228 Journal of the Optics.” Journal of the Optics.
In order to shorten the learning time and improve the recognition rate, the local invariance that the convolution filter should obtain by learning is configured to be easy to learn in advance.
As shown in the example of FIG. 7, the sixth embodiment includes a convolution processing module 710, a rectification + subsampling processing module 720, and a normalization processing module 730.
The convolution processing module 710 performs processing equivalent to the convolution processing module 610.

The rectification + subsampling processing module 720 includes a rectification processing module 722a, a rectification processing module 722b, a rectification processing module 722c, a rectification processing module 722d, a sum calculation processing module 724a, a sum calculation processing module 724d, a subsampling processing module 726a, and a subsampling processing. A module 726d, an averaging processing module 728a, and an averaging processing module 728d are provided.
The rectification processing module 722a is connected to the convolution processing module 710a and the sum calculation processing module 724a. The rectification processing module 722a receives the processing result of the convolution processing module 710a as an input, and performs a process of performing the r-th power on the absolute value of the input (r is a positive integer, preferably, for example, a square of r = 2 The rectification process by the calculation process is performed.
The sum calculation processing module 724a is connected to the rectification processing module 722a, the rectification processing module 722b, and the sub-sampling processing module 726a. The sum calculation processing module 724a adds the processing results of the two rectification processing modules 722 (rectification processing modules 722a and 722b).
The sub-sampling processing module 726a is connected to the sum calculation processing module 724a and the averaging processing module 728a. The sub-sampling processing module 726a performs sub-sampling processing on the processing result by the sum calculation processing module 724a.
The averaging processing module 728a is connected to the sub-sampling processing module 726a, the difference calculation processing module 732, and the sum calculation processing module 734. The averaging processing module 728a performs the process of performing the power of r ′ on the processing result of the sub-sampling processing module 726a (preferably r ′ = 1 / r, for example, the square root process when r = 2). Perform an averaging process.

The normalization processing module 730 includes a difference calculation processing module 732, a sum calculation processing module 734, and a division processing module 736. The normalization processing module 730 extracts the feature amounts of the images or feature maps 100A, 100B, and 100C by performing normalization processing on the processing results of the two averaging processing modules 728. And it outputs to the feature extraction layer of the next stage. Or it outputs to a discriminator. The discriminator identifies the images or feature maps 100A, 100B, and 100C based on the received feature amount.
The difference calculation processing module 732 is connected to the averaging processing module 728a, the averaging processing module 728d, and the division processing module 736.
The sum calculation processing module 734 is connected to the averaging processing module 728a, the averaging processing module 728d, and the division processing module 736.
The division processing module 736 is connected to the difference calculation processing module 732 and the sum calculation processing module 734.
As the normalization processing performed by the normalization processing module 730, the difference between the processing results of the two averaging processing modules 728 (for example, the averaging processing module 728a and the averaging processing module 728d) is used, and the sum of the processing results is used as a divisor. Normalize by dividing. Note that the difference between the processing results is that one processing result is subtracted from the other processing result, and either the averaging processing module 728a or the averaging processing module 728d is one and may be the other. .

なお、ｓは畳込み処理の出力特徴量、ｕは整流及びサブサンプリング処理の出力、ｙは特徴抽出層の出力、ｉ、ｐは特徴量マップのインデックス、ｊ、ｋは特徴量マップ内の２次元座標、ｑ、ｒは畳込みフィルタの２次元座標、ｗは予め与えられる重み付き一般化平均のパラメータである。
正規化処理は、（式１２）の第２式としてもよい。 In this case, (Equation 11) is rewritten as (Equation 12).

Note that s is an output feature value of the convolution process, u is an output of the rectification and sub-sampling process, y is an output of the feature extraction layer, i and p are feature map indices, and j and k are 2 in the feature map. Dimensional coordinates, q and r are two-dimensional coordinates of the convolution filter, and w is a weighted generalized average parameter given in advance.
The normalization process may be the second expression of (Expression 12).

  A hardware configuration example of the image processing apparatus according to the above-described embodiment will be described with reference to FIG. The configuration shown in FIG. 8 is configured by a personal computer (PC), for example, and shows a hardware configuration example including a data reading unit 817 such as a scanner and a data output unit 818 such as a printer.

  The CPU (Central Processing Unit) 801 is the various modules described in the above-described embodiments, that is, the convolution processing module 110, the rectification processing module 120, the normalization processing module 130, the sub-sampling processing module 140, and the feature extraction layer 1 A control unit that executes processing according to a computer program describing an execution sequence of each module such as 210.

  A ROM (Read Only Memory) 802 stores programs used by the CPU 801, operation parameters, and the like. A RAM (Random Access Memory) 803 stores programs used in the execution of the CPU 801, parameters that change as appropriate during the execution, and the like. These are connected to each other via a host bus 804 including a CPU bus.

  The host bus 804 is connected to an external bus 806 such as a PCI (Peripheral Component Interconnect / Interface) bus via a bridge 805.

  A keyboard 808 and a pointing device 809 such as a mouse are input devices operated by an operator. The display 810 includes a liquid crystal display device or a CRT (Cathode Ray Tube), and displays various types of information as text or image information.

  An HDD (Hard Disk Drive) 811 includes a hard disk, drives the hard disk, and records or reproduces a program executed by the CPU 801 and information. The hard disk stores a target image, a recognition result, a learning result, and the like. Further, various computer programs such as various other data processing programs are stored.

  The drive 812 reads out data or a program recorded in a removable recording medium 813 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and reads the data or program as an interface 807 or an external bus 806. , And supplied to the RAM 803 connected via the bridge 805 and the host bus 804. The removable recording medium 813 can also be used as a data recording area similar to a hard disk.

  The connection port 814 is a port for connecting the external connection device 815, and has a connection unit such as USB or IEEE1394. The connection port 814 is connected to the CPU 801 and the like via the interface 807, the external bus 806, the bridge 805, the host bus 804, and the like. The communication unit 816 is connected to a communication line and executes data communication processing with the outside. The data reading unit 817 is a scanner, for example, and executes document reading processing. The data output unit 818 is, for example, a printer, and executes document data output processing.

  Note that the hardware configuration of the image processing apparatus illustrated in FIG. 8 illustrates one configuration example, and the present embodiment is not limited to the configuration illustrated in FIG. 8, and the modules described in the present embodiment are executed. Any configuration is possible. For example, some modules may be configured with dedicated hardware (for example, Application Specific Integrated Circuit (ASIC), etc.), and some modules are in an external system and connected via a communication line In addition, a plurality of systems shown in FIG. 8 may be connected to each other via communication lines so as to cooperate with each other. Further, it may be incorporated in a copying machine, a fax machine, a scanner, a printer, a multifunction machine (an image processing apparatus having any two or more functions of a scanner, a printer, a copying machine, a fax machine, etc.).

  Note that the above-described various embodiments may be combined (for example, adding or replacing a module in one embodiment in another embodiment), and processing contents of each module The technology described in the background art (including the technology described as a reference in the description of “Mode for Carrying Out the Invention”) may be adopted.

The program described above may be provided by being stored in a recording medium, or the program may be provided by communication means. In that case, for example, the above-described program may be regarded as an invention of a “computer-readable recording medium recording the program”.
The “computer-readable recording medium on which a program is recorded” refers to a computer-readable recording medium on which a program is recorded, which is used for program installation, execution, program distribution, and the like.
The recording medium is, for example, a digital versatile disc (DVD), which is a standard established by the DVD Forum, such as “DVD-R, DVD-RW, DVD-RAM,” and DVD + RW. Standard “DVD + R, DVD + RW, etc.”, compact disc (CD), read-only memory (CD-ROM), CD recordable (CD-R), CD rewritable (CD-RW), Blu-ray disc ( Blu-ray Disc (registered trademark), magneto-optical disk (MO), flexible disk (FD), magnetic tape, hard disk, read-only memory (ROM), electrically erasable and rewritable read-only memory (EEPROM (registered trademark)) )), Flash memory, Random access memory (RAM) SD (Secure Digital) memory card and the like.
The program or a part of the program may be recorded on the recording medium for storage or distribution. Also, by communication, for example, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a wired network used for the Internet, an intranet, an extranet, etc., or wireless communication It may be transmitted using a transmission medium such as a network or a combination of these, or may be carried on a carrier wave.
Furthermore, the program may be a part of another program, or may be recorded on a recording medium together with a separate program. Moreover, it may be divided and recorded on a plurality of recording media. Further, it may be recorded in any manner as long as it can be restored, such as compression or encryption.

DESCRIPTION OF SYMBOLS 100 ... Image or feature map 110 ... Convolution processing module 120 ... Rectification processing module 130 ... Normalization processing module 132 ... Difference calculation processing module 134 ... Divisor calculation processing module 136 ... Division processing module 140 ... Subsampling processing module 200 ... Image 210 ... Feature extraction layer 1:
220 ... Feature map A:
230 ... feature extraction layer 2:
240 ... feature map B:
250 ... feature extraction layer 3:
260 ... feature map C:
270 ... Output 300 ... Image or feature map 310 ... Convolution processing module 320 ... Rectification processing module 330 ... Sub-sampling processing module 340 ... Normalization processing module 342 ... Difference calculation processing module 344 ... Divisor calculation processing module 346 ... Division processing module 400 ... image or feature map 410 ... convolution processing module 420 ... rectification processing module 430 ... sub-sampling processing module 440 ... normalization processing module 442 ... difference calculation processing module 444 ... divisor calculation processing module 446 ... division processing module 500 ... image or feature Map 510 ... Convolution processing module 520 ... Rectification processing module 530 ... Sub-sampling processing module 540 ... Normalization processing module 542 ... Difference calculation processing module 54 ... sum calculation processing module 546 ... division processing module 600 ... image or feature map 610 ... convolution processing module 620 ... rectification + subsampling processing module 622 ... rectification processing module 624 ... sum calculation processing module 626 ... subsampling processing module 628 ... average Normalization processing module 630 ... Normalization processing module 632 ... Difference calculation processing module 634 ... Sum calculation processing module 636 ... Division processing module 700 ... Image or feature map 710 ... Convolution processing module 720 ... Rectification + subsampling processing module 722 ... Rectification processing Module 724 ... Sum calculation processing module 726 ... Sub-sampling processing module 728 ... Averaging processing module 730 ... Normalization processing module 732 ... Difference calculation processing module 34 ... Sum calculation processing module 736 ... Division processing module 900 ... Image 910 ... Feature extraction layer 920 ... Feature extraction layer 930 ... Discriminator 1000 ... Image or feature map 1010 ... Convolution processing module 1020 ... Rectification processing module 1030 ... Normalization processing Module 1032 ... Average calculation processing module 1034 ... Difference calculation processing module 1036 ... Standard deviation calculation processing module 1038 ... Division processing module 1040 ... Sub-sampling processing module

Claims (11)

Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of normalization processing means for performing normalization processing on the processing result by the rectification processing means,
A plurality of feature extraction means for extracting the feature quantity of the image by performing a sub-sampling process on the processing result by the normalization processing means;
An identification unit for identifying the image based on the feature amount extracted by the feature extraction unit;
The normalization processing unit normalizes the difference between the processing results of the two rectification processing units by dividing the difference based on the value based on the processing result. Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the rectification processing means;
A plurality of feature extraction means for extracting feature amounts of the image by performing a normalization process on the processing result of the sub-sampling processing means;
An identification unit for identifying the image based on the feature amount extracted by the feature extraction unit;
The normalization processing means normalizes the difference between the processing results of the two sub-sampling processing means by dividing the difference based on the processing results of the two rectification processing means as a divisor. Image processing device. Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the rectification processing means;
A plurality of feature extraction means for extracting feature amounts of the image by performing a normalization process on the processing result of the sub-sampling processing means;
An identification unit for identifying the image based on the feature amount extracted by the feature extraction unit;
The normalization processing unit normalizes the difference between the processing results of the two sub-sampling processing units by dividing a difference based on the processing result as a divisor. Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the rectification processing means;
A plurality of feature extraction means for extracting feature amounts of the image by performing a normalization process on the processing result of the sub-sampling processing means;
An identification unit for identifying the image based on the feature amount extracted by the feature extraction unit;
The normalization processing unit normalizes the difference between the processing results of the two sub-sampling processing units by dividing the difference between the sums of the processing results as a divisor. Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of addition means for adding the processing results of the plurality of rectification processing means;
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the adding means;
A plurality of averaging processing means for performing an averaging process on the processing result of the sub-sampling processing means;
A plurality of feature extraction means for extracting feature values of the image by performing a normalization process on a processing result by the averaging processing means;
An identification unit for identifying the image based on the feature amount extracted by the feature extraction unit;
The processing by the rectification processing means, the addition means, the sub-sampling processing means, and the averaging processing means is a weighted generalized averaging process,
The normalization processing unit normalizes the difference between the processing results of the two averaging processing units by dividing the difference between the sums of the processing results as a divisor. The rectification processing performed by the rectification processing means is processing (r is a positive real number) for performing the r-th power on the absolute value of the input,
The image processing apparatus according to claim 5, wherein the averaging process performed by the averaging processing unit is a process of performing r ′ power. Computer
Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of normalization processing means for performing normalization processing on the processing result by the rectification processing means,
A plurality of feature extraction means for extracting the feature quantity of the image by performing a sub-sampling process on the processing result by the normalization processing means;
Based on the feature amount extracted by the feature extraction means, function as an identification means for identifying the image,
The normalization processing means normalizes the difference between the processing results of the two rectification processing means by dividing the difference based on the value based on the processing results. Computer
Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the rectification processing means;
A plurality of feature extraction means for extracting feature amounts of the image by performing a normalization process on the processing result of the sub-sampling processing means;
Based on the feature amount extracted by the feature extraction means, function as an identification means for identifying the image,
The normalization processing means normalizes the difference between the processing results of the two sub-sampling processing means by dividing the difference based on the processing results of the two rectification processing means as a divisor. Image processing program. Computer
Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the rectification processing means;
A plurality of feature extraction means for extracting feature amounts of the image by performing a normalization process on the processing result of the sub-sampling processing means;
Based on the feature amount extracted by the feature extraction means, function as an identification means for identifying the image,
The normalization processing means normalizes the difference between the processing results of the two sub-sampling processing means by dividing the difference based on the value based on the processing results. Computer
Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the rectification processing means;
A plurality of feature extraction means for extracting feature amounts of the image by performing a normalization process on the processing result of the sub-sampling processing means;
Based on the feature amount extracted by the feature extraction means, function as an identification means for identifying the image,
The normalization processing means normalizes the difference between the processing results of the two sub-sampling processing means by dividing the difference between the sums of the processing results as a divisor. Computer
Image receiving means for receiving images;
A plurality of convolution processing means for performing convolution processing on the image received by the image receiving means;
A plurality of rectification processing means for performing rectification processing on the processing result by the convolution processing means,
A plurality of addition means for adding the processing results of the plurality of rectification processing means;
A plurality of sub-sampling processing means for performing sub-sampling processing on the processing result by the adding means;
A plurality of averaging processing means for performing an averaging process on the processing result of the sub-sampling processing means;
A plurality of feature extraction means for extracting feature values of the image by performing a normalization process on a processing result by the averaging processing means;
Based on the feature amount extracted by the feature extraction means, function as an identification means for identifying the image,
The processing by the rectification processing means, the addition means, the sub-sampling processing means, and the averaging processing means is a weighted generalized averaging process,
The normalization processing means normalizes the difference between the processing results of the two averaging processing means by dividing the difference between the sums of the processing results as a divisor.

Images