A pattern classification method based on multi-scale basic belief assignment negation

Abstract

This invention relates to image processing technology and image pattern classification, specifically providing a pattern classification method based on multi-scale basic belief assignment negation. The method includes: dividing a preprocessed image into feature maps of different scales using a feature pyramid network; convolving the feature maps and outputting a total feature map of fully connected layer features; calculating different distances for features at different scales in the total feature map, and adding point-to-point correlation coefficients and membership degrees to calculate basic belief assignment values; horizontally concatenating the basic belief assignment values ​​to form a basic belief assignment map for each class; defining the negation of basic belief assignments, expressing all basic belief assignments on the basic belief assignment map in negated form, and calculating the accuracy of the model corresponding to the quality function; calculating the negation of the weights corresponding to the quality function of basic belief assignments; and calculating the output class based on the negation of the weights corresponding to the quality function and the negation of the basic belief assignments, thereby achieving efficient identification of image classes and improving identification accuracy.

Description

Technical Field

[0001] This invention relates to image processing technology and image pattern classification, specifically providing a pattern classification method based on multi-scale basic belief assignment negation. Background Technology

[0002] Pattern classification plays a crucial role in image processing and computer vision, and is widely used in image recognition, medical diagnosis, security monitoring, autonomous driving, and other fields. This task involves preprocessing images, extracting image features, and then feeding the features and corresponding labels into a classification model for training. The performance of the trained model is evaluated using accuracy, recall, and F1 score.

[0003] Multi-scale pattern classification methods leverage feature information from images at different resolutions to recognize that features in an 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 from different scales, image content can be understood and classified more comprehensively. This also helps the classifier remain robust to challenges such as scale changes, occlusion, and noise. Furthermore, combining features from different scales allows for more accurate identification and classification of objects in the image.

[0004] Basic belief assignment is a core concept in Dempster-Shafer evidence theory, used to represent the degree of confidence in a hypothesis. In pattern classification, basic belief assignment can be used to handle uncertainty and incomplete information, particularly in image processing and computer vision, where this theory can help improve classification accuracy and robustness. 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. By dividing the image into multiple scales through a feature pyramid network, it can better capture the high-level semantic information and low-level detail information of the image. Combined with deep learning and Dempster-Shafer evidence theory, it can achieve efficient identification of image categories.

[0006] To achieve the above objectives, 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. Preprocess the image data;

[0009] S2. The preprocessed image is divided into feature maps of different scales using a feature pyramid network.

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

[0011] S4. Calculate different distances for features at different scales in the overall feature map, and add the point-to-two sequence correlation coefficient and membership degree to calculate the basic belief assignment value;

[0012] S5. Horizontally stitch together the basic belief assignment values ​​to form a basic belief assignment diagram for each class;

[0013] S6. Define the negation of basic belief assignments, express all basic belief assignments on the basic belief assignment diagram in the form of negation, and calculate the accuracy of the model corresponding to the quality function.

[0014] S7. Based on the accuracy of the model corresponding to the quality function, calculate the negation of the weights corresponding to the quality function of the basic belief assignment;

[0015] S8. Calculate the output category based on the negation of the weights corresponding to the quality function and the negation of the basic belief assignment.

[0016] Preferably, S1 includes:

[0017] The image data is converted from integer format to 32-bit floating-point number and normalized to the range of 0 to 1.

[0018] Preferably, S2 includes:

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

[0020] Preferably, S3 includes:

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

[0022] Preferably, S4 includes:

[0023] Combining Dempster-Shafer evidence theory, the distance between high-scale features and class prototype vectors is calculated using Manhattan distance, while the distance between low-scale features and class prototype vectors is calculated using Euclidean distance. The point bisequence correlation coefficient is added when calculating the distance.

[0024] The basic belief assignment value is calculated by combining distance with Gaussian function and membership degree.

[0025] Preferably, the formula for calculating the allocation value of each basic belief is as follows:

[0026]

[0027] in, A represents i Assume the basic belief assignment of the j-th sample under the g-th quality function. For membership degree, and The scale and width of the basic belief assignment values ​​are controlled separately, where σ is the standard deviation of the distribution. The distance between the feature vector and the class vector of the j-th sample is given by the following formula:

[0028]

[0029] in, For in A i Under the assumption of the g-th quality function, the t-th feature of the j-th sample, Let r be the total number of features, and δ be the feature of the t-th class prototype of the j-th sample under the g-th quality function. gt (A i The calculation formula is as follows:

[0030]

[0031] Where, δ gt (A i ) indicates in A i Assume the absolute value of the correlation coefficient between two sequences for the t-th feature under the g-th quality function. Represented as A i The mean of a continuous variable under the assumption, Indicates not A i The mean of a continuous variable under the assumption, s gt (A i () represents the standard deviation of a continuous variable. Represented as A i Number of samples under the assumption Indicates not A i The number of samples under the assumption, n gt (A i () represents the total number of samples.

[0032] Preferably, all basic belief assignments on the basic belief assignment map are expressed in negative form, and the calculation formula is as follows:

[0033]

[0034] in, A represents i The negation of the fundamental belief assignment of the g-th mass function under the assumption, |A i ∪A k | represents A i With A kThe number of elements in the intersection of |A i ΔA k | represents A i With A k Union minus A i With A k The number of elements in the intersection of the two sets. A represents the assignment of negation values ​​to the fundamental beliefs of the empty set. i ≠A k Φ = {L1, L2, ..., L p}where L p This represents the p-th class.

[0035] Preferably, S7 includes:

[0036] The negation of the accuracy of the model corresponding to the mass function is calculated using the accuracy of the model corresponding to the mass function. The formula is as follows:

[0037] ε g =1-γ g

[0038] Where, γ g ε represents the accuracy of the model corresponding to the g-th quality function. g This indicates a negation of the accuracy of the model corresponding to the g-th quality function;

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

[0040]

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

[0042] Preferably, S8 includes:

[0043] The weighted Euclidean norm based on the negation of the quality function weights and the negation of the basic belief assignment is calculated for each image. The class with the smallest weighted Euclidean norm is selected as the output class. The calculation formula is as follows:

[0044]

[0045] in, w represents the product of the quality function and the square of its negation weights. g This indicates the negation of the weight corresponding to the g-th quality function. Indicates in A i The negation of the basic quality allocation under the assumption;

[0046]

[0047] Where c represents the category index into which the image is ultimately classified, and G represents the total number of quality functions.

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

[0049] First, converting image data from integer to floating-point format provides better numerical stability and reduces quantization errors during subsequent normalization by dividing by 255.0. Second, normalizing the data and scaling pixel values ​​to the range of 0 to 1 helps improve model training efficiency and performance. Third, using a feature pyramid network generates multiple feature maps at different scales, fully utilizing the advantages of features at each scale, including both semantic information from high-level feature maps and detailed information from low-level feature maps. Currently, determining the basic belief assignment remains an unresolved issue. Current construction methods do not consider the reliability of each feature, leading to inaccuracies when handling uncertain information. This invention uses different methods to calculate the basic information assignment for features at different scales. 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. Considering the importance of each feature, a point-to-secondary correlation coefficient is added when calculating the distance, thereby increasing the reliability of the basic belief assignment construction. This also makes the constructed basic information assignment more closely aligned with the features at the corresponding scale and provides a more comprehensive representation of information. Secondly, the negation of basic belief assignment is defined. Current methods rarely explore basic belief assignment from a negation perspective, but the negation of basic belief assignment proposed in this invention offers a novel approach to image classification. Next, considering the differences between different quality functions and their varying reliability, the negation weight for each quality function is defined by subtracting the accuracy obtained from the model constructing each quality function from 1 and normalizing it. Finally, the square root of the square of the negation weight multiplied by the negation value for each feature is obtained, and the weighted Euclidean norm based on the negation of basic belief assignment is taken for each image. The class with the smallest norm is then selected as the output class. This weighting method improves classification accuracy, especially when features have different importance. The weighted Euclidean norm takes into account the importance of different quality functions, and by assigning different weights, it can more accurately measure the similarity between the image and the class prototype. Furthermore, combining the negation of basic belief assignment and the weighted Euclidean norm provides an objective evaluation standard that quantifies the difference between the image and the predefined class, providing a clear numerical basis for classification decisions. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0051] Figure 1 A flowchart illustrating the pattern classification method based on multi-scale basic belief assignment negation provided by this invention;

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

[0053] Figure 3 The flowchart for constructing basic belief assignments in the pattern classification method based on multi-scale basic belief assignment negation provided by this invention;

[0054] Figure 4 This invention defines basic belief assignment negation and image classification in the pattern classification method based on multi-scale basic belief assignment negation. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] This invention provides a pattern classification method based on multi-scale basic belief assignment negation. This method combines basic belief assignment with deep learning to more efficiently and accurately classify images. Figure 1 As shown, it includes:

[0057] S1. Preprocess the image data;

[0058] S2. The preprocessed image is divided into feature maps of different scales using a feature pyramid network.

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

[0060] S4. Calculate different distances for features at different scales in the overall feature map, and add the point-to-two sequence correlation coefficient and membership degree to calculate the basic belief assignment value;

[0061] S5. Horizontally stitch together the basic belief assignment values ​​to form a basic belief assignment diagram for each class;

[0062] S6. Define the negation of basic belief assignments, express all basic belief assignments on the basic belief assignment diagram in the form of negation, and calculate the accuracy of the model corresponding to the quality function.

[0063] S7. Based on the accuracy of the model corresponding to the quality function, calculate the negation of the weights corresponding to the quality function of the basic belief assignment;

[0064] S8. Calculate the output category based on the negation of the weights corresponding to the quality function and the negation of the basic belief assignment.

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

[0066] In step S1, the input image is first acquired, and the array containing the image data is converted to a 32-bit floating-point number using the astype function. Then, the elements are divided by 255.0 to scale the range to between 0 and 1, ensuring that the input features are on the same scale, which helps with model training and gradient descent.

[0067] In step S2, the image at a uniform scale is fed into a feature pyramid network with ResNet-101 as the backbone. The input image is passed through a bottom-up path via a conv1 layer containing 7x7 convolutional layers, followed by BatchNorm and ReLU activation functions, and then a 3x3 max-pooling layer to obtain feature map C1. Feature map C1 passes through layer 1, containing 3 residual blocks, outputting feature map C2 with a downsampling factor of 4. Feature map C2 passes through layer 2, containing 4 residual blocks, outputting feature map C3 with a downsampling factor of 8. Feature map C3 passes through layer 3, containing 23 residual blocks, outputting feature map C4 with a downsampling factor of 16. Feature map C4 passes through layer 4, containing 3 residual blocks, outputting feature map C5 with a downsampling factor of 32. Then, through a top-down path, feature map C5 is passed through a 1x1 convolutional layer, reducing the number of channels from 2048 to 256, resulting in feature map F5. Feature map F5 is upsampled to the same size as feature map C4, then horizontally concatenated with the output of layer 3 using a 1x1 convolution. Finally, a 3x3 convolution is used to eliminate aliasing, resulting in feature map F4. Next, feature map F4 is upsampled and horizontally concatenated with the output of layer 2, followed by a 3x3 convolution to obtain feature map F3. Feature map F3 is then upsampled and horizontally concatenated with the output of layer 1, followed by a 3x3 convolution to obtain feature map F2. Finally, five feature maps of different scales are output: F2, F3, F4, F5, and C5. Each feature map corresponds to a different stage of ResNet, capturing information at different scales, such as image texture, color, and shape.

[0068] In step S3, the low-scale feature maps F5 and C5 obtained in step S2 are processed using the Big-Little Net model to obtain fully connected layer features. Through a multi-branch structure, high accuracy is maintained while reducing computational cost, enabling efficient convolutional fully connected layers. The high-scale feature maps F2, F3, and F4 are processed using the ResNet18 model to obtain fully connected layer features. This model includes multiple convolutional layers and residual blocks, effectively extracting features from the high-scale feature maps. Furthermore, each image is labeled with a category, and each category is extracted separately. Assuming there are p categories, five total feature maps containing fully connected layer features from all images are finally extracted at different scales (F2, F3, F4, F5, and C5) using the corresponding models. f2 P f3 P f4 P f5 and P c5 Each graph contains categories y1, y2, ..., y p .

[0069] In step S4, such as Figure 3 As shown, for the high-scale feature map P obtained in step S3... f2 P f3 P f4 The feature a is calculated using Manhattan distance. p and class prototype vector b p The distance between them, for low-scale feature maps P f5 and P c5 The Euclidean distance is used to calculate the distance between the feature and the class prototype vector. To account for the importance of each feature, feature a is calculated when calculating the distance. p and the real category y p The correlation coefficient between the two sequences is used to increase the reliability of the basic belief assignment construction. Then, the distance is combined with the Gaussian function and membership degree to calculate the basic belief assignment value of the p-th class under the g-th quality function. Ultimately, each category corresponds to a basic belief assignment map containing G quality functions, and the output of the basic belief assignment maps for all categories is m. The basic belief assignment formula for each category is expressed as:

[0070]

[0071] in A represents i Assume that the basic belief assignment of the j-th sample under the g-th quality function is... For membership degree, and Controlling the scale and width of the basic belief assignment values, where σ is the standard deviation of the distribution, The distance between the feature vector and the class vector of the j-th sample is given by the following formula:

[0072]

[0073] in, For in A i Under the assumption that the t-th feature of the j-th sample under the g-th quality function, Let δ be the feature of the t-th class prototype of the j-th sample under the g-th quality function, and r be the total number of features. gt (A i The correlation coefficient () represents the correlation between the basic belief assignment and the class prototype for each feature, and is expressed by the following formula:

[0074]

[0075] Where, δ gt (A i ) indicates in A i Assuming the absolute value of the correlation coefficient between two sequences for the t-th feature under the g-th quality function, Represented as A i The mean of a continuous variable under the assumption, Indicates not A i The mean of a continuous variable under the assumption, s gt (A i () represents the standard deviation of a continuous variable. Represented as A i Number of samples under the assumption Indicates not A i The number of samples under the assumption, n gt (A i () represents the total number of samples.

[0076] In step S5, the basic belief assignments obtained in S4 through different networks at different scales are horizontally stitched together to form a basic belief assignment map, thereby obtaining a complete basic belief assignment map for each category.

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

[0078]

[0079] in, A represents i The negation of the fundamental belief assignment of the g-th mass function under the assumption, |Ai ∪A k | represents A i With A k The number of elements in the intersection of |A i ΔA k | represents A i With A k Union minus A i With A k The number of elements in the intersection of the two sets. A represents the assignment of negation values ​​to the fundamental beliefs of the empty set. i ≠A k Φ = {L1, L2, ..., L p}where L p This represents the p-th class.

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

[0081] ε g =1-γ g

[0082] Where γ g ε represents the accuracy of the model corresponding to the g-th quality function. g This indicates a negation of the accuracy of the model corresponding to the g-th quality function.

[0083]

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

[0085] In step S8, the square of the negation of the weights corresponding to the quality function is multiplied by the negation of the basic belief assignment. Then, the sum of these values ​​for each feature is taken as the square root to obtain the final weighted Euclidean norm based on the negation of the basic belief assignment for each image. The class with the smallest norm is then selected as the output class. The formula is as follows:

[0086]

[0087] in w represents the product of the quality function and the square of its negation weights. g This indicates the negation of the weight corresponding to the g-th quality function. Indicates hypothesis A i The negation of the basic quality allocation.

[0088]

[0089] Where c represents the category index to which the image is ultimately classified, 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 Each sample is ultimately classified into class L. C (x1), L C (x2), ..., L C (xn), L C For y c , indicating that the sample was ultimately classified into class c.

[0090] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A pattern classification method based on multi-scale basic belief assignment negation, characterized in that, Includes the following steps: S1. Preprocess the image data; S2. The preprocessed image is divided into feature maps of different scales using a feature pyramid network. S3. Convolve the feature map and output the total feature map of the fully connected layer features; S4. Calculate different distances for features at different scales in the overall feature map, and add the point-to-two sequence correlation coefficient and membership degree to calculate the basic belief assignment value; S5. Horizontally stitch together the basic belief assignment values ​​to form a basic belief assignment diagram for each class; S6. Define the negation of basic belief assignments, express all basic belief assignments on the basic belief assignment diagram in the form of negation, and calculate the accuracy of the model corresponding to the quality function. S7. Based on the accuracy of the model corresponding to the quality function, calculate the negation of the weights corresponding to the quality function of the basic belief assignment; S8. Calculate the output category based on the negation of the weights corresponding to the quality function and the negation of the basic belief assignment.

2. The pattern classification method based on multi-scale basic belief assignment negation as described in claim 1, characterized in that, S1 includes: The image data is converted from integer format to 32-bit floating-point number and normalized to the range of 0 to 1.

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

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

5. The pattern classification method based on multi-scale basic belief assignment negation as described in claim 1, characterized in that, S4 includes: Combining Dempster-Shafer evidence theory, the distance between high-scale features and class prototype vectors is calculated using Manhattan distance, while the distance between low-scale features and class prototype vectors is calculated using Euclidean distance. The point bisequence correlation coefficient is added when calculating the distance. The basic belief assignment value is calculated by combining distance with Gaussian function and membership degree.

6. The pattern classification method based on multi-scale basic belief assignment negation as described in claim 5, characterized in that, The formula for calculating the basic belief assignment value for each category is as follows: in, A represents i Assume the basic belief assignment of the j-th sample under the g-th quality function. For membership degree, and The scale and width of the basic belief assignment values ​​are controlled separately, where σ is the standard deviation of the distribution. The distance between the feature vector and the class vector of the j-th sample is given by the following formula: in, For in A i Under the assumption of the g-th quality function, the t-th feature of the j-th sample, Let r be the total number of features, and δ be the feature of the t-th class prototype of the j-th sample under the g-th quality function. gt (A i The calculation formula is as follows: Where, δ gt (A i ) indicates in A i Assume the absolute value of the correlation coefficient between two sequences for the t-th feature under the g-th quality function. Represented as A i The mean of a continuous variable under the assumption, Indicates not A i The mean of a continuous variable under the assumption, s gt (A i () represents the standard deviation of a continuous variable. Represented as A i Number of samples under the assumption Indicates not A i The number of samples under the assumption, n gt (A i () represents the total number of samples.

7. The pattern classification method based on multi-scale basic belief assignment negation as described in claim 1, characterized in that, Express all basic belief assignments on the basic belief assignment map in negative form, calculated using the following formula: in, A represents i The negation of the fundamental belief assignment of the g-th mass function under the assumption, |A i ∪A k | represents A i With A k The number of elements in the intersection of |A i ΔA k | represents A i With A k Union minus A i With A k The number of elements in the intersection of the two sets. A represents the assignment of negation values ​​to the fundamental beliefs of the empty set. i ≠A k Φ={L1,L2,…,L p }, L p This represents the p-th category.

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

9. A pattern classification method based on multi-scale basic belief assignment negation as described in claim 1, characterized in that, S8 includes: The weighted Euclidean norm based on the negation of the quality function weights and the negation of the basic belief assignment is calculated for each image. The class with the smallest weighted Euclidean norm is selected as the output class. The calculation formula is as follows: in, w represents the product of the quality function and the square of its negation weights. g This indicates the negation of the weight corresponding to the g-th quality function. Indicates in A i The negation of the basic quality allocation under the assumption; Where c represents the category index into which the image is ultimately classified, and G represents the total number of quality functions.

Images