Adaptive and multi-scale image feature extraction method

Abstract

The invention discloses a self-adaptive and multi-scale image feature extraction method, and relates to the field of image feature extraction, and the method specifically comprises the following steps: S1, carrying out the feature space mapping of an original image; s2, enhancing texture and edge features; s3, carrying out multi-scale feature reconstruction; S4, carrying out dynamic convolution kernel; s5, performing high-dimensional feature decomposition and reconstruction; s6, carrying out feature fusion; s7, carrying out orthogonalization constraint; s8, calculating a loss function; s9, optimizing model parameters; processing the image under different scales; and S10, obtaining an enhanced image. The method has the advantages of multi-scale feature fusion: ensuring that important features can be extracted under different scales by carrying out multi-scale processing on the image and combining a weighted fusion mechanism of the features, and dynamic convolution kernel generation: designing a dynamic convolution kernel generation mechanism, so that the convolution kernel can dynamically adjust the weight according to the input features, thereby improving the robustness of the image. Therefore, the method can better adapt to the local change of the image.

Description

Technical Field

[0001] The present invention relates to the field of image feature extraction, and in particular to an adaptive and multi-scale image feature extraction method. Background Art

[0002] Image feature extraction technology is widely used in many fields such as target recognition, image classification and object tracking. However, the complexity of images in actual application scenarios, such as illumination changes, noise interference, multi-scale features and dynamic changes of local features, brings great challenges to feature extraction.

[0003] To address these issues, traditional convolutional neural networks (CNNs) have made significant progress in feature extraction, but they still have limitations in dealing with illumination changes, detail loss, and feature redundancy. In particular, when an image contains features at multiple scales, a single-scale convolution kernel is difficult to capture all important information. In addition, the model is prone to extracting redundant features, which increases the computational burden and leads to a decrease in generalization performance.

[0004] In order to enhance the adaptability and robustness of the model, multi-scale feature extraction and adaptive feature enhancement techniques have gradually become research hotspots. These technologies can better capture detailed information in the image by dynamically adjusting the convolution kernel and using multi-level feature fusion. At the same time, as the complexity of the model increases, controlling the feature sparsity of the model and reducing feature redundancy have also become important strategies to improve performance.

[0005] In this context, an adaptive and multi-scale feature extraction algorithm is proposed. By introducing an adaptive convolution kernel generation mechanism, multi-scale feature fusion and orthogonal constraints, the performance of the model in complex environments is improved. In addition, by combining sparsity regularization and regularization loss, feature redundancy and overfitting are further reduced, and the generalization ability of the model is enhanced.

[0006] The illumination changes in the real environment will affect the visibility of objects in the image, making it more difficult for the model to recognize them. Different objects and details may appear at different scales, so a single-scale feature extraction method often cannot obtain enough information. As the model extracts more features, some features may be redundant, increasing the computational cost and model complexity. The dynamic changes in local areas in the image require the convolution kernel to be able to adapt flexibly to effectively extract local features.

[0007] In existing research, methods such as multi-scale feature fusion and dynamic convolution kernel generation can improve the feature extraction ability of the model. For example, the multi-scale convolutional neural network (MSCNN) extracts features simultaneously through convolution kernels of different scales; the dynamic convolution kernel method dynamically adjusts the weights of the convolution kernel according to the input to adapt to complex image features. However, most of these methods focus on the robustness of feature extraction, and rarely involve how to control feature redundancy and improve feature independence.

[0008] To overcome these challenges, this paper proposes an adaptive and multi-scale image feature extraction algorithm. Summary of the invention

[0009] The purpose of the present invention is to solve the problems in the prior art and to propose an adaptive and multi-scale image feature extraction method.

[0010] An adaptive and multi-scale image feature extraction method introduces an adaptive convolution kernel generation mechanism, multi-scale feature fusion and orthogonal constraints to improve the performance of the model in complex environments. At the same time, it combines sparsity regularization and regularization loss to further reduce feature redundancy and overfitting. The specific steps include:

[0011] S1: feature space mapping of the original image;

[0012] S2: Enhance texture and edge features;

[0013] S3: Multi-scale feature reconstruction, processing images at different scales;

[0014] S4: dynamic convolution kernel;

[0015] S5: Decomposition and reconstruction of high-dimensional features. Decompose the high-dimensional features into multiple low-rank tensors, and then perform CP decomposition to decompose the high-dimensional tensor into a set of low-dimensional sub-tensors. For each decomposed low-dimensional sub-tensor, apply convolution operations respectively. Each convolution result represents a part of the enhanced features. Finally, all convolution results are weighted fused to obtain the final enhanced features. Through decomposition and reconstruction, the multi-dimensional dependency between features is captured, thereby significantly improving the expressiveness of the model.

[0016] S6: Feature fusion, each convolution result represents a part of the enhanced features, then all convolution results are weighted and fused to obtain the final enhanced features;

[0017] S7: Orthogonalization constraint,The goal of the orthogonalization constraint is to ensure that the extracted features have high independence from each other and reduce redundancy, thereby improving the classification accuracy and enhancing the generalization ability of the model;

[0018] S8: loss function calculation;

[0019] S9: Optimize model parameters; introduce sparsity regularization to encourage the model to extract only a small number of important features, enhance the model's interpretability and prevent overfitting

[0020] S10: Obtain an enhanced image.

[0021] In the above-mentioned adaptive and multi-scale image feature extraction method, in step S1, it is assumed that the input image I (x, y) is a two-dimensional signal, and a feature space mapping function φ (I (x, y)) is defined, and its basic form is:

[0022] φ(I(x,y))=σ(W·I(x,y)+b)

[0023] Where: W is a weight matrix used to adjust the features of the input image; b is the bias; σ(·) is a nonlinear activation function, using the ReLU function, and the mapping is:

[0024] F φ (x, y)=max(0, W·I(x, y)+b)

[0025] Then, the features of the input image are mapped to a non-negative high-dimensional space.

[0026] In the above-mentioned adaptive and multi-scale image feature extraction method, in step S2, the second-order derivative of the input image is introduced Represents the local gradient change, which is specifically calculated as:

[0027]

[0028] By adding this term, enhanced features are obtained:

[0029]

[0030] where γ is a hyperparameter that controls the weight of the gradient boosting term.

[0031] In the above-mentioned adaptive and multi-scale image feature extraction method, when reconstructing the multi-scale features, for the s-th scale, the image is downsampled. s (x, y), and perform weighted summation on the mapping results of each scale:

[0032]

[0033] in: is the feature map at scale s; α s is the weight of the corresponding scale.

[0034] In the above-mentioned adaptive and multi-scale image feature extraction method, in step S4, the generation mechanism of the dynamic convolution kernel specifically includes the following steps: Assuming that a local area feature of the input image is F(x, y), the convolution kernel W(x, y) is adaptively generated along with the input feature, and the core formula for generating the convolution kernel is:

[0035]

[0036] Where W0 is the weight of the initial convolution kernel, which is the weight obtained after pre-training; θ i It is the weight vector of the dynamic adjustment convolution kernel, and its value is determined by the input feature F(x, y);

[0037] The weights of the dynamic convolution kernel are generated through a linear transformation and activation function:

[0038] W(x, y) = σ(W0 + ∑ i，j θ i，j F(x+i,y+j))

[0039] Where: θ i，j is the local weight generated based on the input features; F(x+i, y+j) is the neighborhood feature of the convolution operation; σ(·) is the nonlinear activation function;

[0040] For each input feature region, calculate its corresponding convolution kernel W(x, y) and apply it to the current feature region:

[0041] F conv (x,y)=Σ i，j W(x+i, y+j)·F(x+i, y+j).

[0042] In the above-mentioned adaptive and multi-scale image feature extraction method, in step S6, assuming that the input feature tensor is T(x, y, c), where x and y are spatial dimensions and c is the number of channels, the tensor is decomposed into R sub-tensors using high-order tensor decomposition:

[0043]

[0044] Use CP decomposition to decompose a high-dimensional tensor into a set of low-dimensional sub-tensors. The formula for CP decomposition is:

[0045]

[0046] Among them: a i (x), b i (y) and c i (c) is the decomposed low-dimensional vector; λ i is the weight of each component;

[0047] For each decomposed sub-tensor T i (x, y, c), apply the convolution operation respectively:

[0048] F i (x, y) = Conv(T i (x,y,c)

[0049] Each convolution result F i (x, y) represents a part of the enhanced features. Finally, all convolution results are weighted and fused to obtain the final enhanced features:

[0050]

[0051] In the above-mentioned adaptive and multi-scale image feature extraction method, in step S8, the calculation process is: assuming that the feature vector set extracted by the model is F = {F1, F2, ..., F N}, where F i represents the i-th eigenvector;

[0052] The overall loss function is defined as:

[0053]

[0054] Among them: α, β, and λ are hyperparameters used to adjust the weight of each loss term in the overall loss, is the classification loss, which is used to evaluate the classification performance of the model; It is the orthogonal constraint loss to ensure the independence of features; is the sparsity regularization term, which controls the feature sparsity of the model; is a conventional L2 regularization term used to prevent model overfitting and is often defined as:

[0055]

[0056] Where W i are the parameters of the model;

[0057] The orthogonal constraint loss function is defined as:

[0058]

[0059] Among them: F i ·F j represents the inner product of two eigenvectors; ||F i || and ||F j || are the eigenvectors F i and F j The norm of ; Yes F i and F jThe cosine similarity ranges from -1 to 1.

[0060] In the above-mentioned adaptive and multi-scale image feature extraction method, in step S9, the sparsity regularization is implemented by the L1 norm, and the sparsity constraint of the feature vector F is defined as:

[0061]

[0062] Among them: F ij Denotes the eigenvector F i The jth component of ||F i ||1 is the L1 norm of the feature vector, which is used to encourage sparsity;

[0063] At the same time, cross entropy loss is used as the classification loss to measure the classification performance of the model, which is defined as:

[0064]

[0065] Where: y c is the onehot encoding of the actual category; p c is the model’s predicted probability for class c; C is the number of classes.

[0066] Compared with the prior art, the present invention has the following advantages:

[0067] 1. Multi-scale feature fusion: By performing multi-scale processing on the image and combining the weighted fusion mechanism of the features, it ensures that important features can be extracted at different scales.

[0068] 2. Dynamic convolution kernel generation: A dynamic convolution kernel generation mechanism is designed so that the convolution kernel can dynamically adjust the weights according to the input features, thereby better adapting to local changes in the image.

[0069] 3. Orthogonalization constraint: By imposing orthogonalization constraints on feature vectors, we ensure that the features extracted by the model remain as independent as possible and reduce redundant information.

[0070] 4. Sparsity regularization: Sparsity regularization is introduced, and through the L1 norm constraint, the model is forced to extract only a small number of key features, which enhances the interpretability of the model and prevents overfitting.

[0071] In summary, through the above-mentioned innovative design parts, the present invention makes the method show stronger robustness when dealing with complex scenes such as illumination changes, noise interference and geometric transformations, and at the same time achieves significant improvements in reducing feature redundancy and improving model generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1A schematic diagram of a model in an adaptive and multi-scale image feature extraction method proposed in the present invention. DETAILED DESCRIPTION

[0073] Reference Figure 1 , an adaptive and multi-scale image feature extraction method, first input the original image, assuming that the input original image I (x, y) is a two-dimensional signal, define the feature space mapping function φ (I (x, y)), its basic form is:

[0074] φ(I(x,y))=σ(W·I(x,y)+b)

[0075] Where: W is a weight matrix used to adjust the features of the input image; b is the bias; σ(·) is a nonlinear activation function, using the ReLU function, and the mapping is:

[0076] F φ (x, y)=max(0, W·I(x, y)+b)

[0077] Map the features of the input image to a non-negative high-dimensional space.

[0078] In order to enhance edge and texture features, the second-order derivative of the input image is introduced Represents the local gradient change. Its specific calculation is:

[0079]

[0080] This is used to highlight edges and details. By adding this, we get enhanced features:

[0081]

[0082] where γ is a hyperparameter that controls the weight of the gradient boosting term.

[0083] In multi-scale feature reconstruction, the image is processed at different scales. For the sth scale, the image is downsampled I s (x, y), and perform weighted summation on the mapping results of each scale:

[0084]

[0085] in: is the feature map at scale s; α s is the weight of the corresponding scale, which is adjusted through learning. This formula ensures that the features extracted at multiple scales can be fused, so that important information in the image can be effectively identified at different scales.

[0086] In order to adapt to the dynamic changes of local features of the image, a dynamic convolution kernel generation mechanism is designed so that the convolution kernel weights can be dynamically generated according to the input features.

[0087] Assuming that the local area feature of the input image is F(x, y), the convolution kernel W(x, y) is generated adaptively according to the input feature. The core formula for generating the convolution kernel is:

[0088]

[0089] Where W0 is the weight of the initial convolution kernel, which is the weight obtained after pre-training; θ i It is the weight vector of the convolution kernel that is dynamically adjusted, and its value is determined by the input feature F(x, y).

[0090] The weights of the dynamic convolution kernel are generated through a linear transformation and activation function:

[0091] W(x, y) = σ(W0 + ∑ i，j θ i，j F(x+i,y+j))

[0092] Where: θ i，j is the local weight generated based on the input features; F(x+i, y+j) is the neighborhood feature of the convolution operation; σ(·) is the nonlinear activation function. This mechanism enables the convolution kernel to change dynamically according to the different feature areas of the image, improving the adaptability to complex features.

[0093] For each input feature region, calculate its corresponding convolution kernel W(x, y) and apply it to the current feature region:

[0094] F conv (x, y) = ∑ i，j W(x+i,y+j)·F(x+i,y+j)

[0095] Through this dynamically generated convolution kernel, more effective convolution operations can be performed on features at different positions, improving the robustness of the model in complex situations such as lighting changes and texture changes.

[0096] By decomposing high-dimensional features into multiple low-rank tensors, the nonlinear dependencies between features can be captured, thereby enhancing the feature extraction effect.

[0097] Assume that the input feature tensor is T(x, y, c), where x and y are spatial dimensions and c is the number of channels. Use high-order tensor decomposition to decompose the tensor into R sub-tensors:

[0098]

[0099] Use CP decomposition (CANDECOMP / PARAFAC decomposition) to decompose a high-dimensional tensor into a set of low-dimensional sub-tensors. The formula for CP decomposition is:

[0100]

[0101] Among them: a i (x), b i (y) and c i (c) is the decomposed low-dimensional vector; λ i is the weight of each component.

[0102] For each decomposed sub-tensor T i (x, y, c), apply the convolution operation respectively:

[0103] F i (x, y) = Conv(T i (x,y,c)

[0104] Each convolution result F i (x, y) represents a part of the enhanced features. Finally, all convolution results are weighted fused to obtain the final enhanced features:

[0105]

[0106] Through this decomposition and reconstruction, the multidimensional dependencies between features can be captured, thereby significantly improving the expressiveness of the model.

[0107] In order to optimize the feature extraction ability and classification performance of the model, a loss function based on orthogonalization constraint is proposed. The goal of orthogonalization constraint is to ensure that the extracted features have high independence from each other and reduce redundancy, thereby improving classification accuracy and enhancing the generalization ability of the model.

[0108] Assume that the set of feature vectors extracted by the model is F = {F1, F2, ..., F N}, where F i represents the i-th eigenvector.

[0109] The overall loss function is defined as:

[0110]

[0111] Among them: α, β, and λ are hyperparameters used to adjust the weight of each loss term in the overall loss, is the classification loss, which is used to evaluate the classification performance of the model; It is the orthogonal constraint loss to ensure the independence of features; is the sparsity regularization term, which controls the feature sparsity of the model; is a conventional L2 regularization term used to prevent model overfitting and is often defined as:

[0112]

[0113] Where W i are the parameters of the model.

[0114] In order to keep these eigenvectors as orthogonal as possible and avoid redundant information, the orthogonal constraint loss function is defined as:

[0115]

[0116] Among them: F i ·F j represents the inner product of two eigenvectors; ||F i || and ||F j || are the eigenvectors F i and F j The norm of ; Yes F i and F j The cosine similarity of , ranges from 1 to 1.

[0117] The goal of this project is to make the cosine similarity between feature vectors approach 0, ensuring that the feature vectors are as orthogonal as possible, thereby encouraging the model to extract more complementary and independent features.

[0118] In order to further optimize the model, sparsity regularization is introduced to force the model to extract only a small number of important features, enhance the interpretability of the model and prevent overfitting. Sparsity regularization is usually implemented through the L1 norm. For the sparsity constraint of the feature vector F, it is defined as:

[0119]

[0120] Among them: F ij Denotes the eigenvector F i The jth component of ||F i ||1 is the L1 norm of the feature vector and is used to encourage sparsity.

[0121] CrossEntropy Loss is used as the classification loss to measure the classification performance of the model. It is defined as:

[0122]

[0123] Where: y c is the onehot encoding of the actual category; p c is the model’s predicted probability for class c; C is the number of classes.

[0124] Comparison of experimental results

[0125] To further verify the effectiveness of the algorithm, experiments were conducted on multiple public datasets, including tasks such as image classification, object detection, and image segmentation. The performance of the proposed algorithm was compared with that of MSCNN, Dynamic CNN, and FPN in the following indicators:

[0126] 1. Classification accuracy: measures the performance of the model on image classification tasks.

[0127] 2. Detection accuracy (mAP, mean Average Precision): Evaluates the detection accuracy of the model in the target detection task.

[0128] 3. Feature redundancy: Evaluate the independence between features by calculating the similarity between extracted features.

[0129] 4. Computational efficiency: measures the inference speed and computational complexity of the model.

[0130] 1. Classification task:

[0131] In the classification experiment on the ImageNet dataset, the proposed algorithm has a slight improvement in classification accuracy compared with several other algorithms, especially in the case of multi-scale objects and complex lighting conditions. The introduction of orthogonalization constraints effectively reduces feature redundancy, allowing the model to learn more independent feature representations, thereby improving classification accuracy.

[0132] algorithm Top 1 Accuracy Top 5 Accuracy MSCNN 77.4％ 93.2％ Dynamic CNN 78.1％ 93.5％ FPN 78.5％ 93.8％ Algorithm of the present invention 79.2％ 94.4％

[0133] Second target detection task:

[0134] In the object detection task on the COCO dataset, the proposed algorithm outperforms other algorithms in terms of mAP indicators, especially in the detection of small objects, thanks to the application of multi-scale feature fusion and dynamic convolution kernels. The detection accuracy of the proposed algorithm is 12 percentage points higher than that of FPN and Dynamic CNN.

[0135]

[0136]

[0137] 3. Feature redundancy:

[0138] By calculating the cosine similarity of the extracted feature vectors, the proposed algorithm is significantly lower than other algorithms in feature redundancy, indicating that the orthogonalization constraint plays a significant role in reducing feature redundancy.

[0139] algorithm Average cosine similarity MSCNN 0.36 Dynamic CNN 0.33 FPN 0.30 Algorithm of the present invention 0.25

[0140] 4. Computational efficiency:

[0141] With the same computing resources, the algorithm in this paper has improved computational efficiency due to the introduction of sparsity regularization and dynamic feature fusion mechanism. Especially when processing high-resolution images, the inference time is significantly shorter than FPN.

[0142] algorithm Average inference time (ms) MSCNN 85 Dynamic CNN 80 FPN 110 Algorithm of the present invention 78

[0143] By comparing with similar algorithms such as MSCNN, Dynamic CNN and FPN, the adaptive and multi-scale feature extraction algorithm proposed in this paper has achieved better results in classification and detection tasks, especially in dealing with complex scenes and multi-scale objects. Mechanisms such as orthogonalization constraints, sparsity regularization and dynamic convolution kernel generation make the model not only more independent in feature expression, but also greatly reduce feature redundancy and computational complexity.

[0144] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. An adaptive and multi-scale image feature extraction method, characterized in that: The adaptive convolution kernel generation mechanism, multi-scale feature fusion and orthogonal constraints are introduced to improve the performance of the model in complex environments. At the same time, sparsity regularization and regularization loss are combined to further reduce feature redundancy and overfitting. The specific steps include: S1: feature space mapping of the original image; S2: Enhance texture and edge features; S3: Multi-scale feature reconstruction, processing images at different scales; S4: dynamic convolution kernel; S5: Decomposition and reconstruction of high-dimensional features. Decompose the high-dimensional features into multiple low-rank tensors, and then perform CP decomposition to decompose the high-dimensional tensor into a set of low-dimensional sub-tensors. For each decomposed low-dimensional sub-tensor, apply convolution operations respectively. Each convolution result represents a part of the enhanced features. Finally, all convolution results are weighted fused to obtain the final enhanced features. Through decomposition and reconstruction, the multi-dimensional dependency between features is captured, thereby significantly improving the expressiveness of the model. S6: Feature fusion, each convolution result represents a part of the enhanced features, then all convolution results are weighted and fused to obtain the final enhanced features; S7: Orthogonalization constraint,The goal of the orthogonalization constraint is to ensure that the extracted features have high independence from each other and reduce redundancy, thereby improving the classification accuracy and enhancing the generalization ability of the model; S8: loss function calculation; S9: Optimize model parameters; introduce sparsity regularization to encourage the model to extract only a small number of important features, enhance the model's interpretability and prevent overfitting S10: Obtain an enhanced image.

2. The method for extracting adaptive and multi-scale image features according to claim 1, characterized in that: In step S1, assuming that the input image I(x, y) is a two-dimensional signal, the feature space mapping function φ(I(x, y)) is defined, and its basic form is: φ(I(x,y))=σ(W·I(x,y)+b) Where: W is a weight matrix used to adjust the features of the input image; b is the bias; σ(·) is a nonlinear activation function, using the ReLU function, and the mapping is: F φ (x,y)=max(0,W·I(x,y)+b) Then, the features of the input image are mapped to a non-negative high-dimensional space.

3. The method for adaptive and multi-scale image feature extraction according to claim 1, characterized in that: In step S2, the second-order derivative of the input image is introduced Represents the local gradient change, which is specifically calculated as: By adding this term, enhanced features are obtained: where γ is a hyperparameter that controls the weight of the gradient boosting term.

4. The method for extracting adaptive and multi-scale image features according to claim 1, characterized in that: In step S3, when reconstructing the multi-scale features, for the s-th scale, the image is downsampled I s (x, y), and perform weighted summation on the mapping results of each scale: in: is the feature map at scale s; α s is the weight of the corresponding scale.

5. The method for extracting adaptive and multi-scale image features according to claim 1, characterized in that: In step S4, the generation mechanism of the dynamic convolution kernel specifically includes the following steps: Assuming that a local area feature of the input image is F(x, y), the convolution kernel W(x, y) is adaptively generated with the input feature, and the core formula for generating the convolution kernel is: Where W0 is the weight of the initial convolution kernel, which is the weight obtained after pre-training; θ i It is the weight vector of the dynamic adjustment convolution kernel, and its value is determined by the input feature F(x, y); The weights of the dynamic convolution kernel are generated through a linear transformation and activation function: W(x,y)=σ(W0+∑ i，j θ i，j F(x+i,y+j)) Where: θ i，j is the local weight generated based on the input features; F(x+i, y+j) is the neighborhood feature of the convolution operation; σ(·) is the nonlinear activation function; For each input feature region, calculate its corresponding convolution kernel W(x, y) and apply it to the current feature region: F conv (x,y)=∑ i，j W(x+i,y+i)·F(x+i,y+j)。 6. The method for extracting adaptive and multi-scale image features according to claim 4, characterized in that: In step S6, assuming that the input feature tensor is T(x, y, c), where x and y are spatial dimensions and c is the number of channels, the tensor is decomposed into R sub-tensors using high-order tensor decomposition: Use CP decomposition to decompose a high-dimensional tensor into a set of low-dimensional sub-tensors. The formula for CP decomposition is: Among them: a i (x), b i (y) and c i (c) is the decomposed low-dimensional vector; λ i is the weight of each component; For each decomposed sub-tensor T i (x, y, c), apply the convolution operation respectively: F i (x,y)=Conv(T i (x,y,c)) Each convolution result F i (x, y) represents a part of the enhanced features. Finally, all convolution results are weighted and fused to obtain the final enhanced features:

7. The method for extracting adaptive and multi-scale image features according to claim 4, characterized in that: In step S8, the calculation process is as follows: Assume that the feature vector set extracted by the model is F = {F1, F2, ..., F N }, where F i represents the i-th eigenvector; The overall loss function is defined as: Among them: α, β, and λ are hyperparameters used to adjust the weight of each loss term in the overall loss, is the classification loss, which is used to evaluate the classification performance of the model; It is the orthogonal constraint loss to ensure the independence of features; is the sparsity regularization term, which controls the feature sparsity of the model; is a conventional L2 regularization term used to prevent model overfitting and is often defined as: Where W i are the parameters of the model; The orthogonalized constraint loss function is defined as: Among them: F i ·F j represents the inner product of two eigenvectors; ||F i || and ||F j || are the eigenvectors F i and F j The norm of ; Yes F i and F j The cosine similarity ranges from -1 to 1.

8. The method for extracting adaptive and multi-scale image features according to claim 6, characterized in that: In step S9, sparsity regularization is implemented by L1 norm, and the sparsity constraint of feature vector F is defined as: Among them: F ij Denotes the eigenvector F i The jth component of ||F i ||1 is the L1 norm of the feature vector, which is used to encourage sparsity; At the same time, cross entropy loss is used as the classification loss to measure the classification performance of the model, which is defined as: Where: y c is the onehot encoding of the actual category; p c is the model’s predicted probability for class c; C is the number of classes.

Images