Remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features

Abstract

The present application relates to a remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features, the R, G and B channels of a color image are respectively transformed by Haar wavelet, the first-level wavelet transform feature image is directly taken as the network input layer to extract the deep features of the image, the feature image is secondly transformed by wavelet according to the first-level wavelet transform, the second-level wavelet transform feature image is taken as the input of the wavelet-attention fusion module, and the second-level deep wavelet features are output; the first-level deep wavelet features and the second-level deep wavelet features are fused by broadcast element addition method; then the next-level wavelet features are subjected to layer-by-layer convolution, the feature channels are increased, and then fused with the previous level output; after the deep wavelet features are extracted, two 1x1 convolution layers and a global average pooling layer are respectively passed through; a dropout layer is added after the full connection layer, and finally the class probability is output through the softmax layer for classification prediction. The calculation complexity is reduced, and the classification accuracy is obviously improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a high-resolution remote sensing image classification method based on deep learning, which can be used for target detection and identification of remote sensing images. BACKGROUND

[0002] Remote sensing technology is to receive electromagnetic wave information of various ground targets on the earth's surface from high altitude or outer space through aircraft, artificial satellites, microwave radar and other equipment, so as to observe the resources and environment on the earth, and remote sensing images are thus generated. High-resolution remote sensing image scene classification is the core problem of remote sensing image analysis and interpretation. The spatial resolution of such remote sensing images is generally high, and the ground targets have rich morphological information. When classifying the images, it is not considered which types of ground targets are contained in each image, but the image is taken as a whole, and the ground target categories are determined according to the overall characteristics of the image. At present, the high-resolution remote sensing image scene classification technology has been widely applied in the fields of natural disaster prevention, city management, ground cover type judgment, vegetation mapping, natural environment change detection and the like.

[0003] With the development of deep learning algorithms, especially the convolutional neural network (CNN) model, researchers gradually use deep features to replace low-level features. Compared with traditional algorithms, deep features show strong feature representation ability, which provides a new way to improve the scene classification of remote sensing images. High-resolution remote sensing images have rich features. In order to obtain deep semantic features, convolutional neural network has gradually become one of the core algorithms for solving the classification problem of high-resolution remote sensing images. Christian Szegedy et al. proposed an improved ResNet model, which reduces the number of residual convolution layers to shorten the training time, and integrates features by introducing a 1x1 convolution layer. The experimental results show that the classification effect is better than that of the original structure. Karen Simonyan compared the performance of hand-crafted features and data-driven features in the problem of remote sensing scene classification, and proved that the accuracy and stability of data-driven features are higher, while the operation efficiency of hand-crafted features is higher and the required training samples are less.

[0004] Recent studies have shown that the fusion features from different feature descriptors or different convolution layers have stronger discriminability in scene understanding and greater potential in improving classification accuracy. For example, Professor Zhong Yanfei of Wuhan University et al. extracted six different features of Gabor filter, GLCM, Gaussian wavelet, shape feature, simple radiation feature and SIFT feature, and indexed the remote sensing image according to different combinations of the above features. Piotr Porwik et al. use attention network to locate the multi-scale discriminative region in remote sensing scene image, and then use classification network to fuse the features of the located region. The above researches have achieved remarkable results, which well prove that feature fusion is an effective step for image representation and scene classification. SUMMARY

[0005] TECHNICAL PROBLEMS TO BE SOLVED

[0006] In view of the problem that the deep learning-based high-resolution remote sensing image classification algorithm does not consider all spatial features and does not extract sufficient detail and texture information, the present application provides a remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features to enhance the extraction of texture information by the network and improve the classification accuracy.

[0007] TECHNICAL SCHEME

[0008] A remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features, characterized in that the steps are as follows:

[0009] Step 1: input a color remote sensing image, first perform channel decomposition on the image to obtain three channel components R, G and B of the image; perform Haar wavelet transform on the R, G and B channels of the color image respectively to obtain four components: low-pass low-pass LL, low-pass high-pass LH, high-pass low-pass HL and high-pass high-pass HH; then extract the transformed images of the three channels to obtain the wavelet transform image of the color image;

[0010] Step 2: directly use the wavelet transform feature image of the color image obtained in step 1 as the input layer of the network, wherein the network comprises a 3x3 convolution layer with a step of 1 and a 3x3 convolution layer with a step of 2;

[0011] The initial convolution channel is set to 64, the convolution kernel size is 3x3, the step is 1, and the deep features of the image are extracted; then the activated output is passed through a 3x3 convolution layer with a step of 2, which is equivalent to reducing the height and width to 1 / 2 of the original through downsampling, and the number of channels remains unchanged;

[0012] Step 3: According to the first-level wavelet transform result in step 1, the low-pass low-pass LL, low-pass high-pass LH, high-pass low-pass HL and high-pass high-pass HH feature images are subjected to second-level wavelet transform; each level of multi-level wavelet transform outputs an image, which will be subjected to a down-sampling operation with a factor of 2; when the height and width of the input image are h and w respectively, after j-level wavelet transform, the image will be converted into (h / 2 j )×(w / 2 j );

[0013] Step 4: According to the obtained second-level wavelet transform feature image, it is taken as the input of the wavelet-attention fusion module, and the output of the module is matched through a channel convolution layer, so that the number of channels is equal to the number of channels of the first-level deep wavelet feature;

[0014] The fusion operation rule used by the wavelet-attention fusion module is defined as follows:

[0015] A(Conv(X LL ),S(G(F(Conv(X LH ,X HL ,X HH )))))

[0016] Where, Conv(.) represents convolution operation, the four components after wavelet transform are generated by stacking R, G and B three channels, and the channel number is expanded to c after convolution operation; F(.) represents feature integration, three components are gathered into global detail features by broadcasting element addition; G(.) represents squeezing, data is compressed into a vector of c×1×1 through global average pooling; S(.) represents excitation, the squeezed data is first converted into (c / 4)×1×1 through full connection layer 1, and the activation function is relu; then through full connection layer 2, it is converted into c×1×1, and the activation function is softmax; A(.) represents attention, the excitation weight obtained according to three components is multiplied with the components through broadcast multiplication, and the fused wavelet feature is obtained;

[0017] According to the number of channels of the first-level deep wavelet feature and the second-level deep wavelet feature, the feature map size is equal, and the activation output is fused by broadcasting element addition method, which must ensure that the channel number and the feature map size are equal;

[0018] Step 5: the output of step 4 is outputted through a 3x3 convolution layer with a step of 1 and a 3x3 convolution layer with a step of 2, the convolution module outputs the wavelet fusion feature channel number which is expanded to twice the original; then the second-level wavelet transform result obtained in step 2 is subjected to third-level wavelet transform, step 4 is repeated to obtain deep wavelet fusion features, finally, the obtained deep wavelet fusion features are outputted through a convolution module to expand the wavelet fusion feature channel number to twice the original; the third-level wavelet transform is subjected to fourth-level wavelet transform, step 4 is repeated to obtain final deep wavelet fusion features;

[0019] Step 6: the deep wavelet feature image finally obtained in step 5 is taken as the input of a 1x1 convolution layer, and a deep wavelet feature image after feature integration is outputted, then the wavelet feature image is inputted into a 1x1 convolution layer again, the dimension of the outputted wavelet feature image is halved, finally, an average pooling layer is passed through, and the length and width of the outputted wavelet feature image become 1 / 2 of the original, the feature map size is reduced;

[0020] Step 7: the wavelet feature image finally obtained in step 6 is inputted into a full connection layer, a dropout layer is added after the full connection layer, a part of neurons is not involved in training, and the generalization ability of the network is enhanced; finally, a softmax layer is passed through, and a class probability is outputted for classification prediction.

[0021] Further technical solutions of the application: the definition formula of the wavelet transform is as follows:

[0022]

[0023] i={H,V,D}

[0024] Wherein, (x, y) is the coordinate of the image pixel point, φ j，m，n (x,y) is the scale function of image decomposition, is a wavelet function, i represents the wavelet direction, including the horizontal direction H, the vertical direction V and the diagonal direction D, m and n respectively determine the displacement along the x and y directions, j=1, 2, 3, 4 represents the wavelet decomposition level.

[0025] A computer system, characterized in comprising: one or more processors, a computer readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0026] A computer readable storage medium, characterized in storing computer executable instructions, the instructions being executed to implement the above method.

[0027] Advantages

[0028] The remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features has the following advantages compared with the prior art.

[0029] Firstly, compared with network models such as LeNet-5, AlexNet, VGG16 and ResNet, the model based on LeNet-5 is improved, the maximum pooling layer of the original model is replaced by a convolution layer, so that the network not only has the downsampling function of the maximum pooling layer, but also retains the spatial information of the original image, and the depth of the network is increased to strengthen the feature extraction; in addition, the BN layer and the Dropout layer are added to prevent overfitting and enhance the generalization ability of the network.

[0030] Secondly, the wavelet-attention fusion module is added, which takes the key feature position of the high-frequency component as the attention factor of the low-frequency component, so that the feature extraction during network training pays more attention to the edge contour and texture information of the image. Through deep wavelet feature extraction and attention mechanism, the extraction of deep features is enhanced, and the classification accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0031] The accompanying drawings are included to provide a further understanding of the application, and are incorporated herein and constitute a part of this application. The same reference numerals in the drawings and their corresponding descriptions represent the same or similar components.

[0032] Figure 1 is the implementation flowchart of the application;

[0033] Figure 2 is the wavelet-attention fusion module flowchart;

[0034] Figure 3 is the confusion matrix of two data sets: (a) NWPU RESISC45 data set; (b) AID data set;

[0035] Figure 4 is the training curve of three models of two data sets. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0037] The application provides a remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features, and the method comprises the following steps: respectively performing Haar wavelet transformation on R, G and B channels of a color image, directly taking a first-level wavelet transformation feature image as a network input layer to extract deep features of the image, performing second-level wavelet transformation on the feature image according to the first-level wavelet transformation, taking a second-level wavelet transformation feature image as an input of a wavelet-attention fusion module, and outputting second-level deep wavelet features; fusing the first-level deep wavelet features and the second-level deep wavelet features through a broadcast element addition method; after one-time wavelet feature fusion, the number of output channels of a next layer convolution module is expanded to twice the original number, then the next-level wavelet features are subjected to layer-by-layer convolution to increase the feature channels, and then the features are fused with the features output by previous levels; after the deep wavelet features are extracted, the training features are integrated and the dimension is reduced through two 1*1 convolution layers and a global average pooling layer; a dropout layer is added after a full connection layer to make a part of neurons not participate in training, so that the generalization ability of the network is enhanced. Finally, a class probability is output through a softmax layer to perform classification prediction. The application reduces the calculation complexity, obviously improves the classification accuracy, and can be used for target detection and identification of high-resolution remote sensing images.

[0038] Reference Figure 1 The specific steps of the application are as follows:

[0039] Step 1: inputting a color remote sensing image, and performing Haar wavelet transformation to obtain a first-level wavelet transformation feature image.

[0040] 1.1) inputting a color remote sensing image, and performing channel decomposition on the image to obtain three channel components R, G and B of the image;

[0041] 1.2) respectively performing Haar wavelet transformation on R, G and B channels of the color image to decompose the channels into four components: low-pass low-pass (LL), low-pass high-pass (LH), high-pass low-pass (HL) and high-pass high-pass (HH);

[0042] 1.2.1) the definition formula of the wavelet transformation is as follows:

[0043]

[0044] i={H,V,D}

[0045] wherein (x, y) is the coordinate of an image pixel point, and is a scale function of image decomposition, j,m,n (x,y) is a scale function of image decomposition, is a wavelet function, i represents a wavelet direction, including a horizontal direction H, a vertical direction V and a diagonal direction D, m and n respectively determine the displacement along the x and y directions, j=1, 2, 3, 4 represents the wavelet decomposition level;

[0046] 1.2.2) The four filter definitions of the two-dimensional discrete Haar wavelet transform are as follows:

[0047]

[0048]

[0049] wherein f LL represents extraction of low frequency information, f LH represents extraction of horizontal feature information, f HL represents extraction of vertical feature information, and f HH represents extraction of diagonal information.

[0050] 1.2.3) The operation rules of the two-dimensional discrete Haar wavelet transform are as follows:

[0051] T LL (x,y) = (f LL *T)↓2 = (T(2x-1,2y-1) + T(2x-1,2y) + T(2x,2y-1) + T(2x,2y)) / 4

[0052] T LH (x,y) = (f LH *T)↓2 = (-T(2x-1,2y-1) - T(2x-1,2y) + T(2x,2y-1) + T(2x,2y)) / 4

[0053] T HL (x,y) = (f HL *T)↓2 = (-T(2x-1,2y-1) + T(2x-1,2y) - T(2x,2y-1) + T(2x,2y)) / 4

[0054] T HH (x,y) = (f HH *T)↓2 = (T(2x-1,2y-1) - T(2x-1,2y) - T(2x,2y-1) + T(2x,2y)) / 4

[0055] wherein T represents a color image, * represents convolution operation, and ↓ represents image down-sampling.

[0056] 1.3) The wavelet transform images of the color image are obtained by extracting the transform images of the three channels;

[0057] Step 2: The first-level wavelet transform feature image is taken as the network layer input, the image features are extracted through the convolution module, and the height and width are converted to 1 / 2 of the original through down-sampling.

[0058] 2.1) Set the initial convolutional channel to 64, and the first-level wavelet transform feature image is subjected to a convolution kernel with a size of 3*3 and a step of 1, so that the length and width of the image remain unchanged, and the channel number becomes 64;

[0059] 2.2) Add a BN layer and an activation layer after each convolution operation; the BN layer can reduce the covariance drift in the neural network and accelerate the convergence speed of the network; the activation layer uses Relu as the activation function, which is a commonly used nonlinear activation function in the neural network structure, and its definition is as follows:

[0060]

[0061] 2.3) The activated output is subjected to a convolution kernel with a size of 3*3 and a step of 2, and the height and width of the image are converted to 1 / 2 of the original;

[0062] Step 3, second-level wavelet transform is performed on the feature map.

[0063] According to the first-level wavelet transform, the second-level wavelet transform is performed on the feature map; the input image is subjected to a down-sampling operation with a multiple of 2 at each level of the multi-level wavelet transform. When the height and width of the input image are h and w respectively, after j-level wavelet transform, the image will be converted to (h / 2 j )×(w / 2 j );

[0064] Step 4, referring to Figure 2 , a wavelet-attention fusion module is used to strengthen the extraction of texture information, edge contour and other features.

[0065] According to the obtained second-level wavelet transform feature image, it is used as the input of the wavelet-attention fusion module, and the output of the module is subjected to a channel matching convolution layer so that the channel number is equal to the channel number of the first-level deep wavelet feature.

[0066] 4.1) LL component, LH component, HL component and HH component are subjected to convolution operation respectively, and the channel number is expanded to c after convolution operation, and in Figure 1 , the size of c is 64;

[0067] 4.2) The LH component, HL component and HH component are aggregated into a global detail feature by the broadcast element addition method;

[0068] 4.3) The global detail feature obtained in 4.2) is used as the input of the attention module, and the excitation weight of the three components is obtained, wherein the attention mechanism includes three steps: squeezing, excitation and attention;

[0069] 4.3.1) "squeezing", that is, global average pooling is performed on all feature images, and the formula is:

[0070]

[0071] 4.3.2) "excitation", that is, generating a one-dimensional excitation weight through a full connection layer and an activation layer, the formula is:

[0072] S(z, W) = σ (W2δ (W1z))

[0073] Wherein, δ and σ represent the relu function and the softmax function respectively, W is the dimension of the full connection layer, W1 is 1 / 4 of the input dimension, and W2 is equal to the input dimension;

[0074] 4.3.3) "attention" will be obtained after the excitation excitation weight S and the LL component feature are broadcasted and multiplied:

[0075] A(S, x) = S·x(c, h, w)

[0076] 4.4) combination Figure 2 The fusion operation rule used is defined as follows:

[0077] A(Conv(X LL )(,S(G(F(Conv(X LH ,X HL ,X HH )))))

[0078] Wherein, Conv(.) represents convolution operation, the four components after wavelet transform are generated by stacking R, G and B three channels, and the channel number is expanded to c after convolution operation. F(.) represents feature integration, three components are gathered into global detail features by broadcast element addition. G(.) represents squeezing, through global average pooling, the data is compressed into a vector of c×1×1. S(.) represents excitation, the squeezed data is first converted into (c / 4)×1×1 through a full connection layer 1, and the activation function is relu; then through a full connection layer 2, it is converted into c×1×1, and the activation function is softmax. A(.) represents attention, and the excitation weight obtained according to the three components is broadcasted and multiplied with the components to obtain the fused wavelet feature.

[0079] 4.5) The obtained wavelet feature is matched through a channel convolution layer, the size of the convolution kernel is 1×1, and the step is 1, so that the number of channels is equal to the number of channels of the wavelet feature of the previous level;

[0080] 4.6) The channel of the first level wavelet feature is equal to the channel of the second level wavelet feature, the feature map size is equal, and the activation output is fused by broadcast element addition method;

[0081] Step 5, repeat step 3 and step 4 to extract deep wavelet features.

[0082] After a wavelet feature fusion, the next layer of convolution module output channel number expands to 2 times the original, and then the next level wavelet feature is convolved layer by layer to increase the feature channel, and then fused with the output of the previous level.

[0083] Step 6, after the deep wavelet feature extraction is completed, the training features are integrated and the dimension is reduced by two 1x1 convolution layers and a global average pooling layer.

[0084] 6.1) Through two convolution kernels with a size of 1x1 and a step of 1, the feature integration is realized, and the number of channels is halved to reduce the calculation amount;

[0085] 6.2) Through a global average pooling layer, each channel corresponds to a mean value, and each feature map corresponds to an output feature, which enhances the consistency of the feature map and the category, and does not need parameter optimization to avoid overfitting;

[0086] Step 7, add a dropout layer after the full connection layer, and finally pass through a softmax layer.

[0087] 7.1) Add a dropout layer after the full connection layer to enhance the generalization ability of the network and reduce the overfitting structure of the neural network. The dropout probability of each neural network layer is set to 0.3;

[0088] 7.2) Through the softmax layer, the class probability is output for classification prediction.

[0089] The effect of the present application will be further described in combination with the simulation experiment:

[0090] 1. Experimental conditions

[0091] The experimental simulation environment is MATLAB R2010a, Intel(R) Pentium(R) 2 CPU 2.93GHz, Window XP Professional.

[0092] Two kinds of remote sensing image data sets are selected for the experiment:

[0093] One is the NWPU-RESISC45 data set, which covers 45 kinds of remote sensing scenes such as airplanes, basketball courts and deserts. Each category contains 700 images with a resolution of 256x256, and a total of 31500 remote sensing images.

[0094] The second is the AID data set, which is composed of 30 kinds of remote sensing scenes such as baseball fields, canal bridges and churches. Each category contains 220-420 images with a resolution of 600x600, and a total of 10000 remote sensing images.

[0095] 2. Experimental content

[0096] In order to improve the generalization of the network, the NWPU, AID dataset is expanded 3 times by image enhancement, and the expanded dataset is divided into a training set and a test set in a ratio of 8:2. After preprocessing, the deep wavelet features of the image are extracted, and the network is trained and predicted. In order to objectively evaluate the effectiveness of the model in high-resolution remote sensing image classification, the classification accuracy and the confusion matrix are used as objective evaluation indexes.

[0097] Experiment 1, the experimental objects are NWPU RESISC45 and AID, two publicly available remote sensing image datasets, the batch size of neural network training is set to 32, and the training round is set to 100. During the training process, the learning rate updating strategy is used: lr=lr x 0.5, wherein lr is the learning rate, the training is started as 0.0005, and when the training accuracy does not decrease for 3 times in succession, the learning rate is reduced by half and the training is continued. Repeat 10 groups of experiments, and take the average value of the optimal classification accuracy as the classification result. The experimental results are shown in Table 1.

[0098] Experiment 2, the confusion matrix of the scene classification results for the two datasets is calculated, and the experimental results are shown in Table 2. Figure 3

[0099] Experiment 3, in order to verify the influence of deep feature extraction on the classification effect of high-resolution remote sensing images, the experimental results of the convolutional neural network designed by the application, the wavelet neural network without attention mechanism, LeNet-5, AlexNet, VGG16, ResNet18 and improved ResNet50 are compared, and Table 2 is the overall classification accuracy, Figure 4 and the training curves of three models.

[0100] Table 1 Classification accuracy of different level wavelet transform

[0101] 1(％) 2(％) 3(％) 4(％) 5(％) A 93.13 93.57 94.67 95.45 93.54 B 93.01 93.28 94.49 95.13 93.40

[0102] Table 2 Overall classification accuracy

[0103]

[0104] 3. Analysis of experimental results

[0105] From Table 1, it can be seen that A and B respectively represent NWPU RESISC45 and AID datasets used by the application, when 4-level wavelet decomposition is performed, the classification accuracy obtained is the highest, when 5-level wavelet decomposition is used, the model needs to increase the convolutional layer to match the size of the image, and it can be found from the experiment that this method increases the network depth, but the accuracy decreases. Therefore, the application uses 4-level wavelet decomposition.

[0106] From Figure 3 ​It can be found that for the NWPU RESISC45, the classification accuracy of 42 classes in 45 classes of scenes is more than 90%, and the accuracy of shrubs and glaciers reaches 100%; due to the high similarity between some scenes, such as churches and palaces, the classification accuracy is lower than that of other scenes, but is basically close to 90%. For the AID dataset, the confusing scenes are parks and train stations, which are 86% and 89% respectively; schools are easily misjudged as other scenes containing buildings due to their building characteristics, and the accuracy is the lowest; the accuracy of the remaining scenes is more than 90%. Thus, it is verified that the model has excellent classification effect in most scenes in the two datasets.

[0107] It can be seen from Table 2 and Figure 4 It can be seen from Table 2 and

[0108] From the above analysis, it can be seen that the present application proves the feasibility and effectiveness of extracting wavelet features and using attention mechanism to fuse features in high-resolution remote sensing image classification, and the classification accuracy and computational efficiency are significantly improved.

[0109] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application.

Claims

1. A remote sensing image scene classification method fusing attention mechanism and multi-scale deep texture features, characterized in that The steps are as follows: Step 1: input the color remote sensing image, first perform channel decomposition on the image to obtain three channel components R, G and B of the image; the R, G and B channels of the color image are respectively decomposed into four components: low-pass low-pass LL, low-pass high-pass LH, high-pass low-pass HL and high-pass high-pass HH through Haar wavelet transform, and then the transformed images of the three channels are extracted to obtain the wavelet transform image of the color image; Step 2: the wavelet transform feature image of the color image finally obtained in step 1 is directly used as the input layer of the network, and the network includes a 3x3 convolution layer with a step of 1 and a 3x3 convolution layer with a step of 2; The initial convolution channel is set to 64, the convolution kernel size is 3x3, the step is 1, and the depth features of the image are extracted; then the activated output is passed through a 3x3 convolution layer with a step of 2, which is equivalent to reducing the height and width to 1 / 2 of the original through downsampling, and the number of channels remains unchanged; Step 3: According to the first-level wavelet transform result in step 1, the low-pass low-pass LL, low-pass high-pass LH, high-pass low-pass HL and high-pass high-pass HH feature images are subjected to second-level wavelet transform; each level of multi-level wavelet transform outputs an image, and each input image is subjected to a down-sampling operation with a multiple of 2; when the height and width of the input image are h and w respectively, after j-level wavelet transform, the image is converted into ; Step 4: according to the obtained second-level wavelet transform feature image, the image is taken as the input of the wavelet-attention fusion module, and the output of the module is matched through a channel matching convolution layer, so that the number of channels is equal to that of the first-level depth wavelet feature; The fusion operation rule used by the wavelet-attention fusion module is defined as follows: wherein, denotes convolution operation, and four components after wavelet transform are generated by superposition of three channels of R, G and B, and the number of channels is expanded to c after convolution operation; denotes feature integration, and three components are gathered into global detail features by broadcast element addition; denotes squeezing, and data is compressed into a vector of c×1×1 through global average pooling; denotes excitation, and data after squeezing is first converted into through full connection layer 1, and the activation function is relu; and then converted into through full connection layer 2, and the activation function is softmax; denotes attention, and the excitation weight obtained according to three components is multiplied with the components through broadcast multiplication to obtain fused wavelet features; According to the number of channels of the first-level depth wavelet feature and the second-level depth wavelet feature, the feature map size is equal, and the activated output is fused by broadcast element addition method, which must ensure that the number of channels and the feature map size are equal; Step 5: the output of step 4 is passed through a 3x3 convolution layer with a step of 1 and a 3x3 convolution layer with a step of 2, and the number of wavelet fusion feature channels of the convolution layer output is expanded to twice the original; then the second-level wavelet transform result obtained in step 2 is subjected to third-level wavelet transform, step 4 is repeated, and the depth wavelet fusion feature is obtained; finally, the depth wavelet fusion feature obtained is output through a convolution layer, and the number of wavelet fusion feature channels is expanded to twice the original; the third-level wavelet transform is subjected to fourth-level wavelet transform, step 4 is repeated, and the final depth wavelet fusion feature is obtained; Step 6: the depth wavelet feature image finally obtained in step 5 is taken as the input of a 1x1 convolution layer, and the depth wavelet feature image after feature integration is output; then the wavelet feature image is input into another 1x1 convolution layer, and the dimension of the output wavelet feature image is halved; finally, an average pooling layer is used, and the length and width of the output wavelet feature image are reduced to 1 / 2 of the original, reducing the feature map size; Step 7: the wavelet feature image finally obtained in step 6 is input into a fully connected layer, and a dropout layer is added after the fully connected layer to let a part of neurons not participate in training, thereby enhancing the generalization ability of the network; finally, a softmax layer is used to output class probability for classification prediction. 2.The method of claim 1, wherein the method further comprises: determining a plurality of attention maps based on the fusion attention mechanism; and determining a plurality of multi-scale deep texture features based on the plurality of attention maps. The definition formula of the wavelet transform is as follows: wherein, is the coordinate of the image pixel, is the scaling function of the image decomposition, is the wavelet function, denotes the wavelet direction, including the horizontal direction H, the vertical direction V, the diagonal direction D, and determine the displacement along the and directions, respectively, j = 1, 2, 3, 4 denotes the wavelet decomposition level.

3. A computer system, characterized by It includes: One or more processors, a computer readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method of claim 1.

4. A computer-readable storage medium, characterized in that computer executable instructions stored thereon for, when executed, implementing the method of claim 1.

Images