An image recognition method based on incomplete multi-view clustering

Abstract

The present invention relates to an image recognition method based on incomplete multi-view clustering, which is used to solve the problem of image classification and recognition, and belongs to the field of pattern recognition and machine learning, especially the incomplete multi-view clustering task for image data. The method is analyzed from the following two aspects: 1) multiple graph models describe the same image sample from different angles, maximizing the shared consistency information; 2) each perspective has personalized information that other perspectives do not have, and these inherent differences are effectively used for clustering. Specifically, for an incomplete multi-view image data set described by different features, the method constructs an initial graph model under each perspective, marks the position of the missing image with an indicator matrix, and extracts a correlation matrix to achieve the clustering task of the incomplete multi-view data, thereby effectively improving the accuracy and robustness of image recognition.

Description

Technical Field

[0001] The present invention relates to the fields of pattern recognition and machine learning, and in particular to an incomplete multi-view clustering task for image data. Background Art

[0002] Computer vision is the gateway to artificial intelligence and a continuation of human vision. It enables computers to see, recognize, and think as humans do. With the advent of the big data era, computer vision has gradually become an important research object in many fields such as information science and biomedical engineering.

[0003] The 21st century is an era of information explosion. Images have become an important means for people to obtain, express and transmit information. So far, image classification and recognition is still a difficult point in the field of image processing, because the real world is diverse and complex, and there are many ways to obtain images, which makes different images very different from each other and difficult to describe with a unified method and model. Among them, image recognition analysis technology based on cluster analysis is an extremely important and widely used algorithm in the field of image processing.

[0004] In most computer vision fields such as image processing, image data often comes from different sources or is described by different features. For example, a person can be photographed and recorded by cameras from different angles, and a picture can be described by features such as grayscale, LBP, and HOG. We call image data described by multiple features multi-view data.

[0005] Existing clustering methods for multi-view data assume that all image samples in the image data are complete. However, in reality, some interference factors such as equipment failure and transmission interruption often lead to the missing of some samples in the image data, which brings great challenges to multi-view clustering methods. Therefore, how to reduce the adverse effects of missing images and mine the potential information of clustering models is an issue worthy of attention.

[0006] For an image, different features can describe different information, and there is also a consistent correlation between these different features. For multiple image data from multiple perspectives, the bottleneck problem of multi-view clustering today is how to separate and utilize this shared information and individual information. To solve this bottleneck problem, this research method is analyzed from the following two aspects: 1) Multiple graph models describe the same image sample from different perspectives to maximize their shared consistent information; 2) Each perspective has personalized information that other perspectives do not have, and these inherent differences are effectively used for clustering. Summary of the invention

[0007] In view of the above problems, the present invention proposes a clustering method based on incomplete multi-view data to solve the image classification and recognition problem. For an incomplete multi-view image dataset described by different features, an initial graph model is constructed under each view, the position of the missing image is marked with an indicator matrix, and the correlation matrix is ​​extracted to achieve the clustering task of incomplete multi-view data, effectively improving the accuracy and robustness of image recognition.

[0008] The specific technical solutions are as follows:

[0009] (1) Obtain an incomplete multi-view dataset to be clustered The incomplete multi-view data refers to incomplete image data from which M features are extracted, and there are multiple images under each viewing angle.

[0010] (2) According to the adaptive graph learning method, the dataset Construct the initial graph for each view

[0011] (3) Incomplete multi-image data obtained according to step (2) The incomplete multi-graph clustering method IMGC is used to process it and obtain the consistency graph C and a series of personalized graphs.

[0012] (4) Based on the obtained consistency graph C and a series of personalized graphs Calculate the incidence matrix Among them, S contains both the consistency information shared by multiple graphs and the personalized information unique to each graph. The spectral clustering Ncut operation is performed on S to obtain the clustering result of the image.

[0013] Furthermore, the adaptive graph learning method described in step (2) constructs the initial graph A of the mth perspective (m) The method is as follows:

[0014]

[0015]

[0016] The m-th perspective has a total of N (m) images, belonging to K (m) categories. Among them, They represent the i-th and j-th images under the m-th perspective respectively, express Similarity of two images, initial image of the mth view For the initial graph A (m) , we call is its i,jth node, is the edge connecting these two nodes.

[0017] vector Indicates A (m) The i-th row of , ξ is the regularization parameter.

[0018] Furthermore, the incomplete multi-graph clustering method IMGC described in step (2) is specifically as follows:

[0019] The initial image of M perspectives Substitute the following formula to obtain the consistency graph matrix C and a series of personalized matrices

[0020]

[0021]

[0022] Among them, N (m) and K (m) is the number of input images and the number of categories to which the images belong under the mth viewing angle. i ,· represents the i-th row of the shared consistency similarity graph matrix C, L C represents the Laplacian matrix of C, rank(·) represents the rank of the matrix, c i,j and d i,j (m) Represents C and D respectively (m) The i-th and j-th elements of , δ is the threshold, ⊙ represents the multiplication of the corresponding elements of the matrix, represents the squared F norm, and the trade-off factor γ ranges from {10 -4 , 10 -3 , 10 -2 , 10 -1 , 10 0 , 10 1 , 10 2 , 10 3},

[0023] Indicator Matrix W (m) It is used to mark the loss of input image samples. The values ​​of all elements in its i-th row and i-th column are as follows:

[0024]

[0025] Beneficial Effects

[0026] A clustering method for incomplete multi-view data achieves higher clustering accuracy and recognition ability than other methods, which indicates good prospects in future practical image recognition and classification applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 , oneFlowchart of an image recognition method based on incomplete multi-view clustering. DETAILED DESCRIPTION

[0028] A clustering method for incomplete multi-view data is evaluated on the face dataset UMIST with sample missing rates of 10%, 20%, 30%, 40%, 50% and 60%, respectively. The dataset includes 575 images captured from 20 objects, a total of 20 categories, and 19-48 images per category. We extract the LBP, Gabor, and Gist features of this image dataset as three input view data X (1) , X (2) and X (3) The specific steps include:

[0029] 1) Image preprocessing:

[0030] For incomplete multi-view data X (1) , X (2) and X (3) , according to the adaptive graph learning method, construct the corresponding initial graph data A (1) , A (2) and A (3) , the specific process is as follows:

[0031]

[0032]

[0033] Specifically, for the input image data X at the mth viewing angle (m) , which contains N (m) sample images, belonging to K (m) categories, of which Represent the i-th and j-th samples respectively, and use this adaptive graph method to learn the similarity between two samples Thus forming a similarity graph In this similar figure A (m) We call the sample is the i,jth node, similarity is the edge connecting two nodes. ξ is the regularization parameter.

[0034] 2) Cluster analysis:

[0035] For the incomplete multi-image data obtained The incomplete multi-view clustering method IMGC proposed in this study is used to analyze and process it, and the consistency graph matrix C and a series of personalized matrices are obtained.

[0036]

[0037]

[0038] Among them, N (m) and K (m) is the number of input images and the number of categories to which the images belong under the mth viewing angle. i ,· represents the i-th row of the shared consistency similarity graph matrix C, L C represents the Laplacian matrix of C, rank(·) represents the rank of the matrix, c i,j and d i,j (m) Represents C and D respectively (m) The i-th and j-th elements of , δ is the threshold, ⊙ represents the multiplication of the corresponding elements of the matrix, represents the squared F norm, and the trade-off factor γ ranges from {10 -4 , 10 -3 , 10 -2 , 10 -1 , 10 0 , 10 1 , 10 2 , 10 3},

[0039] Indicator Matrix W (m) It is used to mark the loss of input image samples. The values ​​of all elements in its i-th row and i-th column are as follows:

[0040]

[0041] For the obtained C and a series of D (m) , let the incidence matrix Then S contains both the consistent information shared by multiple graphs and the personalized information unique to each graph. While ensuring the global structure, it also explores the local differences. The spectral classification Ncut operation is performed on S to obtain the clustering result.

Claims

1. An image recognition method based on incomplete multi-view clustering, characterized in that: (1) Obtain an incomplete multi-view dataset to be clustered Incomplete multi-view data refers to incomplete image data from which M features are extracted, with multiple images at each view. (2) According to the adaptive graph learning method, the dataset Construct the initial graph for each view (3) The initial image of each viewpoint obtained in step (2) The incomplete multi-graph clustering method IMGC is used to process it and obtain the consistency graph C and a series of personalized graphs. (4) Based on the obtained consistency graph C and a series of personalized graphs Calculate the incidence matrix Among them, S contains both the consistency information shared by multiple graphs and the personalized information unique to each graph. The spectral clustering Ncut operation is performed on S to obtain the clustering result of the image; The adaptive graph learning method described in step (2) constructs the initial graph A of the mth view (m) The method is as follows: The m-th perspective has a total of N (m) images, belonging to K (m) categories, among which They represent the i-th and j-th images under the m-th perspective respectively, express Similarity of two images, initial image of the mth view 1≤i,j≤N (m) ; vector Indicates A (m) The i-th row of , ξ is the regularization parameter; The incomplete multi-graph clustering method IMGC described in step (3) is as follows: The initial image of M perspectives Substitute the following formula to obtain the consistency graph matrix C and a series of personalized matrices Among them, N (m) and K (m) is the number of input images and the number of categories to which the images belong under the mth viewing angle; c i，· represents the i-th row of the shared consistency similarity graph matrix C, L C represents the Laplacian matrix of C, rank(·) represents the rank of the matrix, c i，j and d i，j (m) Represents C and D respectively (m) The i-th and j-th elements of , δ is the threshold, ⊙ represents the multiplication of the corresponding elements of the matrix, represents the squared F norm, and the trade-off factor γ ranges from {10 -4 , 10 -3 , 10 -2 , 10 -1 , 10 0 , 10 1 , 10 2 , 10 3 }, Indicator Matrix W (m) It is used to mark the loss of input image samples. The values ​​of all elements in its i-th row and i-th column are as follows:

Images