JP6349477B1 - Similarity determination program - Google Patents

Abstract

A similarity between data strings is determined using a CNN function. A similarity determination program according to the present invention replaces a parameter string with luminance image data, and uses a result obtained by extracting a feature amount as an image of the luminance image data by a CNN to determine the similarity between the parameter strings. judge. [Selection] Figure 4

Description

  The present invention relates to a technique for determining a similarity between parameter strings.

  Currently, a convolutional neural network (convolution neural network, CNN) has been studied as a technique for determining the similarity between images. CNN is attracting attention as a technique that can effectively learn the feature amount of an image. As a method for causing the CNN to learn an image, there are supervised learning in which an input image and an output result are learned in pairs, and unsupervised learning in which only the input image is learned.

  The following Patent Document 1 describes a technique related to CNN. This document realizes that when there are a plurality of objects having similar elements even if they are different objects, an object having the elements is accurately and efficiently extracted. The object similarity determination method according to an embodiment uses one or more computers using a convolutional neural network (CNN) including one or more convolution layers and a fully connected layer. A step of extracting a plurality of feature amounts from each of the plurality of objects to the one or more computers, and the plurality of feature amounts from each of the plurality of objects, A step of extracting an output value of all connected layers after one or a plurality of convolution layers of the convolution neural network (CNN), and a conversion process in which the output value of the all connected layers is a range within a predetermined range. And performing a step of extracting the converted output value and a step of determining the similarity of the object based on the converted output value. Is disclosed (see summary).

JP 2017-207947 A

  Various image feature amounts have been proposed in the past. When an image is learned by a learning device, it is common to extract the feature amount of the image in advance and input the feature amount vector to the learning device for learning. On the other hand, when the CNN learns an image, it is common to input the image data itself into the CNN without extracting image feature amounts in advance. This has an advantage that the step of extracting the feature amount of the image in advance can be omitted, and the learning result does not depend on the design of the feature amount.

  On the other hand, the CNN holds the learned feature quantity on the intermediate layer, and the feature quantity is used internally by the CNN itself. Therefore, the feature quantity learned by the CNN is a black box and is not assumed to be used outside the CNN. On the other hand, in order to visualize the results learned by the CNN, attempts have been made to represent the convolution filter inside the CNN as an image.

  The above features of CNN are useful in determining the similarity between image data. On the other hand, regarding data strings other than image data, there is a need for searching for other data similar to certain data in various situations. Under such a need, the present inventor has attempted to determine the similarity between data strings using the function of CNN.

  The similarity determination program according to the present invention determines the similarity between parameter strings by using a result obtained by replacing a parameter string with luminance image data and extracting a feature amount of the luminance image data as an image by CNN.

  According to the similarity determination program according to the present invention, it is possible to search for another similar parameter string without performing an operation of designing and extracting a feature amount in advance for an arbitrary parameter string.

1 is a configuration diagram of a similarity determination device 100 according to Embodiment 1. FIG. It is a specific example of a parameter string. It is a conceptual diagram which shows a mode that a parameter sequence is normalized. It is an example of the luminance image which the luminance image generation part 111 produces | generates.

  FIG. 1 is a configuration diagram of a similarity determination apparatus 100 according to Embodiment 1 of the present invention. The similarity determination device 100 is a device that determines the similarity between parameter strings using the function of a CNN (convolutional neural network) 200. The similarity determination device 100 includes a CPU (Central Processing Unit) 110 and a storage device 120. The similarity determination apparatus 100 and the CNN 200 can be connected via an appropriate network, or the similarity determination apparatus 100 itself can implement the CNN 200.

  The CPU 110 includes a luminance image generation unit 111, a feature amount acquisition unit 112, and a similarity calculation unit 113. Details of these functional units will be described later. These functional units can be configured by using hardware such as a circuit device in which these functions are mounted, or can be configured by causing the arithmetic unit to execute software in which these functions are mounted. In the following, it is assumed that the software is configured as shown in FIG. For convenience of description, each functional unit may be described as an operation subject, but it is the CPU 110 that actually executes these functional units.

  The storage device 120 stores, as data, a parameter string for which the similarity determination device 100 determines the similarity. For example, when searching for the data string 1 to data string 10 that is most similar to the target data string, the data strings 1 to 10 can be stored in the storage device 120 in advance.

  CNN 200 is a neural network typically constructed to determine the similarity between image data. A learning result such as a weight between each neuron of the CNN 200 can be configured on a computer different from the similarity determination device 100, or the learning result is stored on the similarity determination device 100 (for example, on the storage device 120). Can be configured on the similarity determination device 100.

  FIG. 2 is a specific example of the parameter string. The parameter string is a data string having a plurality of parameter values. For example, the parameter string data1 has parameter values of param1 = 0.51, param2 = 10.1,. The number of parameters is not necessarily equal between the parameter strings.

  FIG. 3 is a conceptual diagram showing how the parameter string is normalized. In the parameter row illustrated in FIG. 2, param3 has an extremely large value compared to other parameter values. If such a parameter sequence is compared as it is, the influence of param3 becomes larger than other parameter values, and the parameter sequence cannot be appropriately compared. Therefore, the luminance image generation unit 111 aligns the scales by normalizing the parameter values. The right figure of FIG. 3 is a conceptual diagram of the parameter values after normalization. Here, for convenience of explanation, only param1 to param3 are shown.

  The nth parameter string is expressed as data (n) = {param1 (n), param2 (n),... paramN (n)}. The luminance image generation unit 111 normalizes the parameter value using, for example, the following formula.

data (n) ′ =
{
(Param1 (n) -Min (param1)) / Max (param1),
(Param2 (n) -Min (param2)) / Max (param2),
...
(ParamN (n) -Min (paramN)) / Max (paramN),
}

  FIG. 4 is an example of a luminance image generated by the luminance image generation unit 111. The parameter string after normalization by the above formula is composed of parameter values having substantially the same scale. By regarding this parameter value as the luminance value of the pixel, it is possible to generate luminance image data composed of luminance values as illustrated in FIG. If the parameter value has a value after the decimal point, for example, the luminance value may be scaled so that the maximum value is 255.

  FIG. 4 illustrates luminance image data having 8 × 8 = 64 pixel values. It is not always necessary to generate a square luminance image, and luminance image data having an arbitrary number of vertical and horizontal pixels can be generated. However, since extremely flat luminance image data is formed with characteristic amounts described later being biased, it is generally considered desirable to generate square luminance image data. Depending on the number of parameter values, a square pixel matrix may not be formed. In this case, insufficient pixels may be padded with appropriate default values. For example, the luminance value = 0 (black) can be padded.

  The luminance image data shown in FIG. 4 is derived from the parameter sequence illustrated in FIG. 2, but it is considered that the feature amount as the image can be extracted as long as it is image data. Therefore, the feature amount acquisition unit 112 inputs the luminance image data generated by the luminance image generation unit 111 to the CNN 200, and extracts the feature amount as an image of each luminance image data by extracting the feature amount.

  For example, when it is desired to specify a parameter sequence most similar to data1 from other parameter sequences, the feature amount acquisition unit 112 acquires the feature amount by inputting the luminance image data of data1 to dataN to the CNN 200, and acquires the feature amount. By specifying the luminance image data having the feature quantity closest to the feature quantity, the parameter string most similar to data1 can be specified.

  The similarity calculation unit 113 calculates the similarity between the feature amounts of each luminance image data acquired by the feature amount acquisition unit 112. As a method for calculating the similarity, for example, the distance between feature vectors can be used as the similarity. In this case, the smaller the distance between vectors, the higher the similarity between the feature quantities. Alternatively, if the feature quantity is expressed as a multidimensional array, the inner product of the array can be used as the similarity. Other similar methods may be used to obtain the similarity between feature quantities.

  The similarity calculation unit 113 may output the calculated similarity itself, or in addition to this, may output a result of ranking the parameter strings according to the height of the similarity. Other suitable output formats may be used. The output result may be stored as data in the storage device 120 or may be output via an output device such as a display provided in the similarity determination device 100.

<Modification of the present invention>
In the above embodiment, the luminance image data once created and the feature amount thereof may be stored in an appropriate storage unit such as a database. When obtaining the luminance image data of the new parameter sequence and the similarity between them, the luminance image data or the feature amount stored in advance in the storage unit may be referred to.

  In the above embodiment, it has been described that the parameter string is replaced with the luminance value. This typically assumes the process of replacing each parameter included in the numerical parameter string with a luminance value, but if it can be replaced with a luminance value, the parameter string other than the numerical value may be replaced with luminance image data. it can. For example, even a character can be handled as a numerical value by replacing it with a character code. It may be replaced with a numerical value by other appropriate methods.

  In the above embodiment, the luminance image generation unit 111 to the similarity calculation unit 113 are described as three modules, but two or more of these can be implemented as one module.

100: Similarity determination device 110: CPU
111: luminance image generation unit 112: feature amount acquisition unit 113: similarity calculation unit 120: storage device 200: CNN

Claims (3)

A similarity determination program for causing a computer to execute a process for determining a similarity between parameter strings, wherein the computer
A luminance image generation step of converting the first parameter string into luminance image data by replacing each parameter included in the first parameter string with a luminance value of a pixel;
Inputting the luminance image data to a neural network that extracts a feature amount of the image data;
Obtaining a first feature quantity as image data of the luminance image data from an intermediate layer of the neural network;
Calculating a similarity between the first parameter sequence and the second parameter sequence by comparing the first feature amount with a second feature value acquired from the neural network for a second parameter sequence;
Was executed,
In the luminance image generation step, the computer
A similarity determination program characterized by causing a step of complementing an insufficient pixel to be executed when the number of parameters included in the first parameter row is smaller than the number of pixels of the luminance image data . The similarity determination program according to claim 1, wherein in the luminance image generation step, the computer is caused to execute a step of normalizing each parameter included in the first parameter sequence. The similarity determination program according to claim 1, wherein, in the luminance image generation step, the computer generates the luminance image data having a square shape.

Images