Pattern classification method based on multi-scale basic belief distribution negative

Abstract

The invention relates to the field of image processing technology and image mode classification, and particularly provides a mode classification method based on multi-scale basic belief allocation negation, which comprises the following steps: dividing a preprocessed image into feature maps of different scales through a feature pyramid network; carrying out convolution on the feature map and outputting a total feature map of full connection layer features; different distances are calculated for different scale features in the total feature map, and a point double-sequence correlation coefficient and a membership degree are added to calculate a basic belief distribution value; transversely splicing the basic belief distribution values into a basic belief distribution graph of each class; defining negative of basic belief distribution, expressing all basic belief distribution on the basic belief distribution graph in a negative form, and calculating the accuracy of the model corresponding to the quality function; calculating the negative of the weight corresponding to the quality function of the basic belief distribution; and calculating an output category based on the negative of the weight corresponding to the quality function and the negative of the basic belief distribution, thereby realizing the efficient recognition of the image category and improving the recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology and image pattern classification, and specifically provides a pattern classification method based on multi-scale basic belief allocation negation. Background Art

[0002] Pattern classification plays a vital role in the field of image processing and computer vision, and is widely used in image recognition, medical diagnosis, security monitoring, autonomous driving, etc. This task preprocesses the image, extracts image features, and then puts the features and labels corresponding to the image into the classification model for training. The performance of the trained model is judged by accuracy, recall, and F1 value.

[0003] The multi-scale pattern classification method uses the feature information of the image at different resolutions to realize that the features in the image may exhibit different characteristics at different scales, for example, low frequencies correspond to global features of the image, while high frequencies correspond to local details. By combining this information at different scales, the image content can be understood and classified more comprehensively, and it can also help the classifier remain robust in the face of challenges such as scale changes, occlusion and noise. Combining features at different scales can more accurately identify and classify objects in the image.

[0004] Basic belief assignment is a core concept in Dempster-Shafer evidence theory, which is used to express the degree of trust in a hypothesis. In pattern classification, basic belief assignment can be used to deal with uncertainty and incomplete information, especially in the fields of image processing and computer vision. This theory can help improve the accuracy and robustness of classification. Summary of the invention

[0005] In view of this, the present invention designs a pattern classification method based on multi-scale basic belief assignment negation, which divides the image into multiple scales through a feature pyramid network to better capture the high-level semantic information and low-level detail information of the image. Combining deep learning with Dempster-Shafer evidence theory, it can achieve efficient recognition of image categories.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A pattern classification method based on multi-scale basic belief assignment negation includes the following steps:

[0008] S1, preprocessing the image data;

[0009] S2, divide the preprocessed image into feature maps of different scales through the feature pyramid network;

[0010] S3, convolve the feature map and output the total feature map of the fully connected layer features;

[0011] S4, calculating different distances for different scale features in the total feature map, and adding point bi-serial correlation coefficient and membership degree to calculate the basic belief distribution value;

[0012] S5, horizontally splicing the basic belief distribution values ​​into a basic belief distribution map for each class;

[0013] S6. define the negation of basic belief allocation, express all basic belief allocations on the basic belief allocation graph in the form of negation, and calculate the accuracy of the model corresponding to the quality function;

[0014] S7, calculating the negation of the weight corresponding to the quality function assigned to the basic belief based on the accuracy of the quality function corresponding model;

[0015] S8. The output category of negation calculation based on the negation of the corresponding weight of the quality function and the assignment of the negation of the basic belief.

[0016] Preferably, the S1 comprises:

[0017] Convert image data from integer format to 32-bit floating point numbers and normalize them to the range 0 to 1.

[0018] Preferably, S2 includes:

[0019] The normalized image is input into the feature pyramid network with ResNet-101 as the backbone network, and 5 feature maps of different scales are output, including 2 low-scale feature maps and 3 high-scale feature maps.

[0020] Preferably, S3 includes:

[0021] The Big-Little Net model is used to calculate the fully connected layer features for low-scale feature maps, and the ResNet18 model is used to calculate the fully connected layer features for high-scale feature maps.

[0022] Preferably, S4 includes:

[0023] Combined with Dempster-Shafer evidence theory, the distance between the feature and the class prototype vector is calculated by Manhattan distance for high-scale features, and the distance between the feature and the class prototype vector is calculated by Euclidean distance for low-scale features, and the point biserial correlation coefficient is added when calculating the distance;

[0024] The distance is combined with the Gaussian function and the membership degree to calculate the basic belief distribution value.

[0025] Preferably, the calculation formula for each basic belief allocation value is:

[0026]

[0027] in, Indicates A i Assume that the basic belief distribution of the jth sample under the gth quality function is, is the degree of membership, and respectively control the scale and width of the basic belief distribution value, σ is the standard deviation of the distribution, is the distance between the feature vector and the category vector of the jth sample, and the formula is as follows:

[0028]

[0029] in, For A i The tth feature of the jth sample under the gth quality function under the assumption that is the feature of the t-th class prototype of the j-th sample under the g-th quality function, r is the total number of features, δ gt (A i ) is calculated as follows:

[0030]

[0031] Among them, δ gt (A i ) means in A i Assume that the absolute value of the point-wise biserial correlation coefficient of the t-th feature under the g-th quality function, Represented as A i The mean of a continuous variable under the assumption, Indicates non-A i The mean of the continuous variable under assumption, s gt (A i ) is the standard deviation of continuous variables, Represented as A i The number of samples under the assumption, Indicates non-A i The number of samples under the assumption, n gt (A i ) is the total number of samples.

[0032] Preferably, all basic belief distributions on the basic belief distribution graph are expressed in a negative form, and the calculation formula is:

[0033]

[0034] in, Indicates A i The negation of the basic belief distribution of the g-th mass function under the assumption that |A i ∪A k | indicates A i With A kThe number of elements in the intersection of |A i ΔA k | indicates A i With A k The union of minus A i With A k The number of elements in the intersection of The basic belief that represents the empty set is assigned a negative value, A i ≠A k ,Φ={L1,L2,…,L p} where L p represents the pth class.

[0035] Preferably, the S7 includes:

[0036] The negation of the accuracy of the model corresponding to the quality function is calculated by the accuracy of the model corresponding to the quality function. The calculation formula is:

[0037] ε g =1-γ g

[0038] Among them, γ g represents the accuracy of the model corresponding to the g-th quality function, ε g represents the negation of the accuracy of the model corresponding to the g-th quality function;

[0039] The negation of the weight corresponding to the quality function is calculated by negating the accuracy of the model corresponding to the quality function. The calculation formula is:

[0040]

[0041] Among them, w g represents the negation of the weight corresponding to the g-th quality function, and G represents the total number of quality functions.

[0042] Preferably, the S8 includes:

[0043] Based on the negation of the weight of the quality function and the negation of the basic belief distribution, the weighted Euclidean norm of the final negation based on the basic belief distribution corresponding to each image is calculated, and the category with the smallest weighted Euclidean norm is selected as the output category. The calculation formula is:

[0044]

[0045] in, represents the product of the quality function and the square of its negative weight, w g represents the negation of the weight corresponding to the g-th mass function, Indicates that in A i The negation of the basic mass distribution under the assumption of ;

[0046]

[0047] Among them, c represents the category subscript into which the image is finally classified, and G represents the total number of quality functions.

[0048] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following achievable effects:

[0049] First, converting the image data from integer format to floating point format can provide better numerical stability and reduce the quantization error when dividing by 255.0 for normalization. Then, normalizing the data to scale the pixel value to a range of 0 to 1 helps improve the model training efficiency and performance. Secondly, using a feature pyramid network to generate multiple feature maps of different scales can make full use of the feature advantages of each scale, including both high-level feature map semantic information and low-level feature map detail information. How to determine the basic belief allocation is still an unresolved issue. The reliability of each feature is not considered in the construction method, which makes it inaccurate when processing uncertainty information. In the present invention, different scale features are used to calculate the basic information allocation in different ways. For high-scale features, Manhattan distance is used to calculate the distance between the feature and the class prototype vector. For low-scale features, Euclidean distance is used to calculate the distance between the feature and the class prototype vector. Considering the importance of each feature, the point bi-serial correlation coefficient is added when calculating the distance, thereby increasing the reliability of the basic belief allocation construction, and this makes the constructed basic information allocation more suitable for the features of the corresponding scale and more comprehensive in representing information. Secondly, the negation of basic belief allocation is defined. Nowadays, few methods explore basic belief allocation from the perspective of negation. The negation of basic belief allocation proposed in the present invention uses basic belief allocation for image classification from a new perspective. Then, considering that there will be differences between different quality functions and the reliability of each quality function is different, the negation weight of each quality function is defined by subtracting 1 from the accuracy obtained in each model of building the quality function and normalizing it. Finally, the square of the weight of the basic belief allocation negation is multiplied by the basic belief allocation negation value, and then the value under each feature is added and squared to obtain the weighted Euclidean norm of the basic belief allocation negation corresponding to each picture, and then the category with the smallest norm is selected as the output category. This weighted method can improve the accuracy of classification, especially when the features have different importance. The weighted Euclidean norm can take into account the importance of different quality functions. By assigning different weights, the similarity between the picture and the category prototype can be more accurately measured, and the combination of the basic belief allocation negation and the weighted Euclidean norm provides an objective evaluation standard, which can quantify the difference between the picture and the predefined category and provide a clear numerical basis for classification decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0051] Figure 1 A flow chart of a pattern classification method based on multi-scale basic belief assignment negation provided by the present invention;

[0052] Figure 2 The architecture diagram of the pattern classification method based on multi-scale basic belief assignment negation provided by the present invention;

[0053] Figure 3 A flowchart of constructing basic belief allocation in the pattern classification method based on multi-scale basic belief allocation negation provided by the present invention;

[0054] Figure 4 The present invention provides a method for pattern classification based on multi-scale basic belief assignment negation, which defines basic belief assignment negation and image classification. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The embodiment of the present invention provides a pattern classification method based on multi-scale basic belief allocation negation, which can more efficiently and accurately classify images by combining basic belief allocation with deep learning. Figure 1 As shown, including:

[0057] S1, preprocessing the image data;

[0058] S2, divide the preprocessed image into feature maps of different scales through the feature pyramid network;

[0059] S3, convolve the feature map and output the total feature map of the fully connected layer features;

[0060] S4, calculating different distances for different scale features in the total feature map, and adding point bi-serial correlation coefficient and membership degree to calculate the basic belief distribution value;

[0061] S5, horizontally splicing the basic belief distribution values ​​into a basic belief distribution map for each class;

[0062] S6. define the negation of basic belief allocation, express all basic belief allocations on the basic belief allocation graph in the form of negation, and calculate the accuracy of the model corresponding to the quality function;

[0063] S7, calculating the negation of the weight corresponding to the quality function assigned to the basic belief based on the accuracy of the quality function corresponding model;

[0064] S8. The output category of negation calculation based on the negation of the corresponding weight of the quality function and the assignment of the negation of the basic belief.

[0065] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Figure 2 shown.

[0066] In step S1, we first obtain the input image, convert the array containing the image data into 32-bit floating point numbers using the astype function, and then divide the elements by 255.0 to scale the range to between 0 and 1. This ensures that the input features are at the same scale, which helps with model training and gradient descent.

[0067] In step S2, the uniform-scale image is put into the feature pyramid network with ResNet-101 as the backbone network. The input image is passed through the conv1 layer containing a 7x7 convolution layer through a bottom-up path, and then through the BatchNorm and ReLU activation functions, and then through a 3x3 maximum pooling layer to obtain feature map C1. Feature map C1 passes through layer1, contains 3 residual blocks, and outputs feature map C2 with a downsampling factor of 4. Feature map C2 passes through layer2, contains 4 residual blocks, and outputs feature map C3 with a downsampling factor of 8. Feature map C3 passes through layer3, contains 23 residual blocks, and outputs feature map C4 with a downsampling factor of 16. Feature map C4 passes through layer4, contains 3 residual blocks, and outputs feature map C5 with a downsampling factor of 32. Then, through a top-down path, feature map C5 passes through a 1x1 convolution layer to reduce the number of channels from 2048 to 256, and obtains feature map F5. The size of feature map F5 is upsampled to the same size as feature map C4, and then it is horizontally connected with the output of layer3 using 1x1 convolution, and finally the aliasing effect is eliminated by 3x3 convolution to obtain feature map F4. Feature map F4 is then upsampled and horizontally connected with the output of layer2, and then through 3x3 convolution to obtain feature map F3. Feature map F3 is then upsampled and horizontally connected with the output of layer1, and then through 3x3 convolution to obtain feature map F2. Finally, 5 feature maps of different scales are output as F2, F3, F4, F5 and C5. Each feature map corresponds to a different stage of ResNet, and can capture information of different scales such as image texture, color and shape.

[0068] In step S3, the Big-Little Net model is used to obtain the fully connected layer features of the low-scale feature maps F5 and C5 obtained in step S2. Through the multi-branch structure, the convolutional full connection can be efficiently performed while maintaining high accuracy and reducing the amount of calculation. The ResNet18 model is used to obtain the fully connected layer features of the high-scale feature maps F2, F3, and F4. The model includes multiple convolutional layers and residual blocks, which can effectively extract features from high-scale feature maps, and classify each image, and each category is separately proposed. Assuming that there are p categories, 5 total feature maps containing the fully connected layer features of all images obtained by the corresponding model at different scales F2, F3, F4, F5, and C5 are finally extracted. They are P f2 , P f3 , P f4 , P f5 and P c5 , each image contains categories y1, y2, ..., y p .

[0069] In step S4, if Figure 3 As shown, for the high-scale feature map P obtained in step S3 f2 , P f3 , P f4 The Manhattan distance is used to calculate the feature a p and the class prototype vector b p The distance between them, for the low-scale feature map P f5 and P c5 The feature uses the Euclidean distance to calculate the distance between the feature and the class prototype vector. Considering the importance of each feature, the feature a is calculated when calculating the distance. p and the true category y p The point bi-serial correlation coefficient between them is used to increase the reliability of the basic belief allocation construction, and then the distance is combined with the Gaussian function and the membership degree to calculate the basic belief allocation value of the pth category under the gth quality function. Finally, each category corresponds to a basic belief distribution map containing G quality functions, and the basic belief distribution map of all categories is output as m. The basic belief distribution formula for each category is expressed as:

[0070]

[0071] in Indicates A i Assume that the basic belief distribution of the jth sample under the gth quality function is, is the degree of membership, and controls the scale and width of the basic belief distribution values, σ is the standard deviation of the distribution, is the distance between the feature vector and the category vector of the jth sample, and the formula is as follows:

[0072]

[0073] in, For A i Under the assumption that, the tth feature of the jth sample under the gth quality function is, is the feature of the t-th class prototype of the j-th sample under the g-th quality function, r is the total number of features, where δ gt (A i ) is the point biserial correlation coefficient, which indicates the correlation between the basic belief distribution under each feature and the class prototype. The formula is as follows:

[0074]

[0075] Among them, δ gt (A i ) means in A i Assume that the absolute value of the point-wise biserial correlation coefficient of the t-th feature under the g-th quality function is, Represented as A i The mean of a continuous variable under the assumption, Indicates non-A i The mean of the continuous variable under assumption, s gt (A i ) is the standard deviation of continuous variables, Represented as A i The number of samples under the assumption, Indicates non-A i The number of samples under the assumption, n gt (A i ) is the total number of samples.

[0076] In step S5, the basic belief distributions obtained in S4 at different scales calculated by different networks are horizontally spliced ​​into a basic belief distribution map to obtain a complete basic belief distribution map for each category.

[0077] In step S6, the basic belief distribution negation is defined. The basic belief distribution value calculated in step S5 is multiplied by the symmetric difference between the hypothesis and other hypotheses under the same quality function and divided by the union of the hypothesis and other hypotheses, and then the obtained negation value is normalized. The formula is as follows:

[0078]

[0079] in, Indicates A i The negation of the basic belief distribution of the g-th mass function under the assumption that |Ai ∪A k | indicates A i With A k The number of elements in the intersection of |A i ΔA k | indicates A i With A k The union of minus A i With A k The number of elements in the intersection of The basic belief that represents the empty set is assigned a negative value, A i ≠A k ,Φ={L1,L2,…,L p} where L p represents the pth class.

[0080] like Figure 4 As shown, in step S7, the negation of the weight corresponding to each quality function is defined by subtracting the accuracy of the model corresponding to the quality function from 1 and normalizing it, which is used for the subsequent weighted Euclidean norm calculation. The formula is as follows:

[0081] ε g =1-γ g

[0082] where γ g represents the accuracy of the model corresponding to the g-th quality function, ε g represents the negation of the accuracy of the model corresponding to the g-th quality function,

[0083]

[0084] where w g represents the negation of the weight corresponding to the g-th quality function, and G represents the total number of quality functions.

[0085] In step S8, the square of the negation of the weight corresponding to the quality function is multiplied by the negation of the basic belief distribution, and then the value of each feature is added and squared to obtain the final weighted Euclidean norm of the negation of the basic belief distribution corresponding to each image, and then the category with the smallest norm is selected as the output category. The formula is as follows:

[0086]

[0087] in represents the product of the quality function and the square of its negative weight, w g represents the negation of the weight corresponding to the g-th mass function, Indicates hypothesis A i Negation of the basic mass distribution.

[0088]

[0089] Where c represents the category subscript that the image is finally divided into, let sample x n The corresponding original category is y p , then the samples are x1, x2, …, x n and categories y1,y2,…,y p , the final class that each sample is divided into is L C (x1), L C (x2),…,L C (xn), L C for y c , indicating that the sample is finally classified into the cth category.

[0090] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A pattern classification method based on multi-scale basic belief assignment negation, characterized in that: The steps include: S1, preprocessing the image data; S2, divide the preprocessed image into feature maps of different scales through the feature pyramid network; S3, convolve the feature map and output the total feature map of the fully connected layer features; S4, calculating different distances for different scale features in the total feature map, and adding point bi-serial correlation coefficient and membership degree to calculate the basic belief distribution value; S5, horizontally splicing the basic belief distribution values ​​into a basic belief distribution map for each class; S6. define the negation of basic belief allocation, express all basic belief allocations on the basic belief allocation graph in the form of negation, and calculate the accuracy of the model corresponding to the quality function; S7, calculating the negation of the weight corresponding to the quality function assigned to the basic belief based on the accuracy of the quality function corresponding model; S8. The output category of negation calculation based on the negation of the corresponding weight of the quality function and the assignment of the negation of the basic belief.

2. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 1, characterized in that: The S1 includes: Convert image data from integer format to 32-bit floating point numbers and normalize them to the range 0 to 1.

3. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 1, characterized in that: The S2 includes: The normalized image is input into the feature pyramid network with ResNet-101 as the backbone network, and 5 feature maps of different scales are output, including 2 low-scale feature maps and 3 high-scale feature maps.

4. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 3, characterized in that: The S3 includes: The Big-Little Net model is used to calculate the fully connected layer features for low-scale feature maps, and the ResNet18 model is used to calculate the fully connected layer features for high-scale feature maps.

5. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 1, characterized in that: The S4 includes: Combined with Dempster-Shafer evidence theory, the distance between the feature and the class prototype vector is calculated by Manhattan distance for high-scale features, and the distance between the feature and the class prototype vector is calculated by Euclidean distance for low-scale features, and the point biserial correlation coefficient is added when calculating the distance; The distance is combined with the Gaussian function and the membership degree to calculate the basic belief distribution value.

6. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 5, characterized in that: The calculation formula for the basic belief distribution value of each category is: in, Indicates A i Assume that the basic belief distribution of the jth sample under the gth quality function is, is the degree of membership, and respectively control the scale and width of the basic belief distribution value, σ is the standard deviation of the distribution, is the distance between the feature vector and the category vector of the jth sample, and the formula is as follows: in, For A i The tth feature of the jth sample under the gth quality function under the assumption that is the feature of the t-th class prototype of the j-th sample under the g-th quality function, r is the total number of features, δ gt (A i ) is calculated as follows: Among them, δ gt (A i ) means in A i Assume that the absolute value of the point-wise biserial correlation coefficient of the t-th feature under the g-th quality function, Indicated as A i The mean of a continuous variable under the assumption, Indicates non-A i The mean of the continuous variable under assumption, s gt (A i ) is the standard deviation of continuous variables, Indicated as A i The number of samples under the assumption, Indicates non-A i The number of samples under the assumption, n gt (A i ) is the total number of samples.

7. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 1, characterized in that: Express all basic belief distributions on the basic belief distribution diagram in a negative form, and the calculation formula is: in, Indicates A i The negation of the basic belief distribution of the g-th mass function under the assumption that |A i ∪A k | indicates A i With A k The number of elements in the intersection of |A i ΔA k | indicates A i With A k The union of minus A i With A k The number of elements in the intersection of The basic belief that represents the empty set is assigned a negative value, A i ≠A k ,Φ={L1,L2,…,L p }, L p represents the p-th category.

8. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 1, characterized in that: The S7 includes: The negation of the accuracy of the model corresponding to the quality function is calculated by the accuracy of the model corresponding to the quality function. The calculation formula is: e g =1-c g Among them, γ g represents the accuracy of the model corresponding to the g-th quality function, ε g represents the negation of the accuracy of the model corresponding to the g-th quality function; The negation of the weight corresponding to the quality function is calculated by negating the accuracy of the model corresponding to the quality function. The calculation formula is: Among them, w g represents the negation of the weight corresponding to the g-th mass function, and G represents the total number of mass functions.

9. A method for pattern classification based on multi-scale basic belief assignment negation as claimed in claim 1, characterized in that: The S8 includes: Based on the negation of the weight of the quality function and the negation of the basic belief distribution, the weighted Euclidean norm of the final negation based on the basic belief distribution corresponding to each image is calculated, and the category with the smallest weighted Euclidean norm is selected as the output category. The calculation formula is: in, represents the product of the quality function and the square of its negative weight, w g represents the negation of the weight corresponding to the g-th mass function, Indicates that in A i The negation of the basic mass distribution under the assumption of ; Among them, c represents the category subscript into which the image is finally classified, and G represents the total number of quality functions.

Images