Image target recognition method based on convolutional neural network

Abstract

The purpose of the present invention is to provide an image target recognition method based on the idea of ​​convolutional neural network, characterized in that: a local target structure balance matrix is ​​obtained by performing Hadamard product with a pre-constructed change structure balance matrix and a local structure balance matrix in an image, and the number of recognition parameters is increased to improve the recognition accuracy. The present invention uses CNN to extract the local initial structure balance matrix, and performs Hadamad product with the pre-constructed change structure balance matrix to obtain the local target structure balance matrix. The prior art processes the image as a whole without considering the local part, and finally obtains only 3 feature parameters. Now, CNN can be used to obtain multiple groups of local recognition parameter vector groups, which increases the number of recognition parameters and improves the accuracy of image target recognition.

Description

Technical Field

[0001] The present invention relates to the field of machine vision, and in particular to an image target recognition method based on the idea of ​​convolutional neural network. Background Art

[0002] Machine vision is a comprehensive technology, including image processing, mechanical engineering technology, control, electric light source lighting, optical imaging, sensors, analog and digital video technology, computer software and hardware technology (image enhancement and analysis algorithms, image cards, I / O cards, etc.). With the rapid development of technologies such as image processing and pattern recognition, the development of machine vision has been greatly promoted. The development of vision is inseparable from image recognition. Image recognition is an important part of machine vision technology. Accuracy and speed are two important indicators for measuring target image recognition. At present, there are many image recognition methods, such as key point-based image recognition methods, texture-based image recognition methods, model-based image recognition methods, KL-based image recognition methods, geometric feature-based image recognition methods, neural network-based image recognition methods, edge contour-based image recognition methods, and image recognition methods combining complex networks and watershed algorithms. The target recognition method based on structural balance network applies the structural balance network to the field of image target recognition. By combining the contour-based image recognition method with the structural balance network method, the grayscale image preprocessing based on the structural balance network is performed on the image to be recognized, and the initial structural balance matrix of the image to be recognized is obtained. Then, the Hadamard product is performed with the constructed change structural balance matrix to obtain the target structural balance matrix. The characteristic parameters of the target structural balance matrix are extracted and then recognition is performed based on the characteristic parameters. Summary of the invention

[0003] The purpose of the present invention is to provide an image target recognition method based on the idea of ​​convolutional neural network, which can increase the number of recognition parameters and thus improve the recognition accuracy.

[0004] The technical solution of the present invention is: an image target recognition method based on the concept of convolutional neural network, characterized in that: a local target structure balance matrix is ​​obtained by performing Hadamard product of a pre-constructed change structure balance matrix and a local structure balance matrix in the image, and the number of recognition parameters is increased to improve the recognition accuracy.

[0005] The specific steps include:

[0006] Construct a local initial structural equilibrium matrix;

[0007] Image transformation based on Hadamard transform:

[0008] Image feature parameters are extracted based on the classification properties of the structural balance network;

[0009] The image to be identified is identified according to the image feature parameters.

[0010] The “constructing a local initial structural equilibrium matrix” includes:

[0011] The set of all N-order structural equilibrium matrices is denoted as F N ; Let P be F N A mapping on, that is: P:F N →F N ; A 3*3 matrix is ​​used as a pre-constructed local initial structure balance matrix to obtain K=(N-2)2 local initial structure balance matrices.

[0012] The “image transformation based on Hadamard transformation” includes: setting X=(x ij ) is a constructed N-order change structure balance matrix, Y∈F N is the local initial structural balance matrix, Z represents the structural balance matrix after Hadamard transformation, then the mapping Z = XoY is a structural balance matrix transformation relationship, where "o" represents the Hadamard product between matrices; Y is the local initial structural balance matrix; if X = (x ij ) is an N-order binary structure balance matrix, Z = XoY can be expressed as:

[0013] Z ij =sign(x ij )y ij

[0014] Among them, sign represents the sign function, which is defined as follows:

[0015]

[0016] The "extracting image feature parameters according to the classification properties of the structural balance network" includes: using a 3*3 matrix as the changing structural balance matrix and performing Hadamard product with the initial structural balance matrix, and each local initial structural balance matrix obtains three recognition parameters. If there are P local initial structural balance matrices that meet the requirements in the initial structural balance matrix, there will be P*3 feature recognition parameters after the Hadamard product.

[0017] The “identifying the image to be identified according to image feature parameters” includes: identifying the target image using P*3 feature recognition parameters obtained by “extracting image feature parameters according to the classification properties of the structural balance network”.

[0018] The beneficial effects of the present invention are as follows: in the prior art, the image target recognition parameters based on the structural balance network are relatively few, and the recognition accuracy is not very high. The present invention proposes a new convolutional neural network based on which CNN is used to extract the local initial structural balance matrix, and the Hadamad product is performed with the pre-constructed change structural balance matrix to obtain the local target structural balance matrix. The prior art processes the image as a whole without considering the local part, and finally obtains only 3 feature parameters. Now, CNN can be used to obtain multiple groups of local recognition parameter vector groups, which increases the number of recognition parameters and improves the accuracy of image target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of extracting recognition parameter sets for image segmentation DETAILED DESCRIPTION

[0020] The present application is further described below in conjunction with the accompanying drawings.

[0021] An image target recognition method based on the idea of ​​convolutional neural network is characterized by: obtaining a local target structure balance matrix by performing Hadamard product between a pre-constructed change structure balance matrix and a local structure balance matrix in an image, thereby increasing the number of recognition parameters and improving recognition accuracy.

[0022] Specifically, the technical solution of the present invention is divided into the following four steps:

[0023] 1. Construct the local initial structural equilibrium matrix:

[0024] The set of all N-order structural equilibrium matrices is denoted as F N (a subset of the set of symmetric matrices). Let P be F N A mapping on , namely:

[0025] P:F N →F N

[0026] Adopting the idea of ​​CNN, a 3*3 matrix is ​​used as the pre-constructed local initial structure balance matrix

[0027] , for a square matrix of order N, K = (N-2) 2 (N≥3) local initial structure equilibrium matrices.

[0028] 2. Image transformation based on Hadamard transform:

[0029] Hadamard transformation: Let X = (x ij ) is a constructed N-order structural equilibrium matrix, Y∈F Nis the local initial structural balance matrix, Z represents the structural balance matrix after Hadamard transformation, then the mapping Z = XoY is a structural balance matrix transformation relationship, where "o" represents the Hadamard transformation between matrices.

[0030] (Hadamard) product, defined as:

[0031] Z ij =(XoY) ij =x ij y ij

[0032] By z ik z kj z ji =(x ik x kj x ji )(y ik y kj y ji )>0, we can get the above transformation matrix Z is a structural balance matrix.

[0033] Where Z = XoY is called Hadamard structural equilibrium matrix transformation, Y is the local initial structural equilibrium matrix, if X = (x ij ) is an N-order binary structure balance matrix (the matrix elements are only 1 and -1), then the Hadamard transform Z = XoY can be expressed as:

[0034] Z ij =sign(x ij )y ij

[0035] Among them, sign represents the sign function, which is defined as follows:

[0036]

[0037] Using the binary structural balance matrix Hadamard transform is equivalent to using the element sign of the binary matrix X to change the sign of the corresponding element in the matrix Y to obtain the transformation image Z. Using the Hadamard transformation theorem, we know that different local initial structural balance matrices have different Hadamard transformation images. For different images, there will be corresponding Hadamard transformation images. Different images can be distinguished according to different transformation images. Using the basic theorem of structural balance network, we know that all nodes in the structural balance network corresponding to the Hadamard transformation image Z can be divided into two categories. Nodes in the same category have positive connections, and nodes belonging to different categories have negative connections. The positive elements in the matrix Z represent the connection strength between two nodes in the same category, and the negative elements represent the connection strength between two nodes in different categories.

[0038] 3. Image feature parameter extraction based on Hadamard transform

[0039] like Figure 1 Because after the binary structural balance matrix and the structural balance matrix of the original image undergo Hadamard transformation (multiplication of corresponding positions of similar matrices), the elements are only positive and negative. The positive element represents the strength of the connection relationship between two nodes in the same circle of friends, and the negative element represents the connection relationship between two nodes in different circles of friends.

[0040] The average value of the positive connection strength of each category in Z is α1 and α2, and the average value of all negative elements in Z is β. In this way, each local initial structure balance matrix obtains three recognition parameters. Using the idea of ​​CNN, a 3*3 matrix is ​​used as the pre-constructed change structure balance matrix to perform Hadamard product with the initial structure balance matrix corresponding to the pixel points in the target image. Each local initial structure balance matrix obtains three recognition parameters. If there are P local initial structure balance matrices that meet the requirements in the initial structure balance matrix, there are P*3 recognition parameters after the Hadamard product.

[0041] 4. Identify the image to be identified based on the feature parameters

[0042] The target image can be identified by using the P*3 feature parameters obtained above, namely the average values ​​of positive connection strength α1 and α2 and the average value of negative connection strength β.

[0043] The present invention adopts the idea of ​​CNN, performs Hadamad product on the local initial structure balance matrix and the pre-built change structure balance matrix, and obtains the local target structure balance matrix.

[0044] The above description is only a preferred embodiment of the invention and is not intended to limit the invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the invention should be included in the protection scope of the invention.

Claims

1. An image target recognition method based on the idea of ​​convolutional neural network, characterized by: The local target structure balance matrix is ​​obtained by performing Hadamard product between the pre-constructed change structure balance matrix and the local structure balance matrix in the image, so as to increase the number of recognition parameters and improve the recognition accuracy. The image target recognition method based on the convolutional neural network idea comprises the following steps: Construct a local initial structural equilibrium matrix; Image transformation based on Hadamard transform: Image feature parameters are extracted based on the classification properties of the structural balance network; Identify the image to be identified according to the image feature parameters; The constructing of the local initial structural equilibrium matrix comprises: The set of all N-order structural equilibrium matrices is called F N ; Let P be F N A mapping on, that is: P:F N →F N ; Using a 3*3 matrix as the pre-constructed local initial structure balance matrix, we get K = (N-2) 2 A local initial structure equilibrium matrix; where N ≥ 3; The image feature parameter extraction according to the classification property of the structural balance network includes: A 3*3 matrix is ​​used as the change structure balance matrix and the initial structure balance matrix to do Hadamard product. Each local initial structure balance matrix obtains three identification parameters. If there are P local initial structure balance matrices that meet the requirements in the initial structure balance matrix, there are P*3 characteristic identification parameters after doing Hadamard product. The identifying the image to be identified according to the image feature parameters comprises: The target image is identified using the P*3 feature recognition parameters obtained by extracting image feature parameters based on the classification properties of the structural balance network.

2. The image target recognition method based on the convolutional neural network concept according to claim 1 is characterized in that: The image transformation based on Hadamard transformation includes: Let X = (x ij ) is a constructed N-order change structure balance matrix, Y∈F N is the local initial structural balance matrix, Z represents the structural balance matrix after Hadamard transformation, then the mapping is a structural balance matrix transformation relationship, where represents the Hadamard product between matrices; Y is the local initial structure equilibrium matrix; If X=(x ij ) is an N-order binary structure balance matrix, It is expressed as: Z ij =sign(x ij )y ij Among them, sign represents the sign function, which is defined as follows:

Images