Image classification method based on Shannon block attention mechanism multi-scale contour wave network

Abstract

This invention relates to an image classification method based on a multi-scale contourlet network using the Shannon block attention mechanism, comprising: Step 1: acquiring the original remote sensing image and performing standardized preprocessing on the original remote sensing image to obtain the remote sensing image to be tested; Step 2: inputting the remote sensing image to be tested into a trained classification network model to obtain the scene classification result of the original remote sensing image; wherein, the classification network model adopts a multi-scale contourlet network based on the Shannon block attention mechanism. The image classification method of this invention utilizes a multi-scale contourlet network based on the Shannon block attention mechanism to achieve scene classification of remote sensing images. This network, through the advantages of convolutional neural networks in the spatial domain and Contourlet feature learning in the frequency domain, can achieve multi-scale and multi-directional feature extraction of remote sensing images, thereby improving classification accuracy; secondly, Contourlet pooling effectively preserves the edge and texture details of the image, which also improves the classification effect of the remote sensing image to a certain extent.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an image classification method based on a multi-scale contour wave network with Shannon block attention mechanism. Background Technology

[0002] In recent years, with the rapid development of remote sensing technology and sensor systems, remote sensing image data has been emerging continuously. Remote sensing image scene classification is an important means of remote sensing image analysis and interpretation, and it has wide applications in fields such as urban planning, land resource management, and military reconnaissance.

[0003] Remote sensing images are diverse, structurally complex, and rich in information, including both natural features such as deserts, rivers, and oceans, and man-made features such as overpasses, residential areas, and airports. The exceptional complexity of information and the intricate spatial layout of scenes in remote sensing images pose a significant challenge to scene classification and recognition.

[0004] In recent years, Convolutional Neural Networks (CNNs) have demonstrated superior performance in remote sensing image classification, becoming the mainstream classification method. While CNNs can improve classification accuracy to some extent, their lack of ability to extract high-frequency features from images can easily lead to the loss of feature information in remote sensing images. This results in a failure to preserve edge and texture details, ultimately hindering the classification accuracy and impacting the overall effectiveness of remote sensing image classification. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides an image classification method based on a multi-scale contour wave network using a Shannon block attention mechanism. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides an image classification method based on a Shannon block attention mechanism using a multi-scale contour wave network, comprising:

[0007] Step 1: Acquire the original remote sensing image, and perform standardization preprocessing on the original remote sensing image to obtain the remote sensing image to be tested;

[0008] Step 2: Input the remote sensing image to be tested into the trained classification network model to obtain the scene classification result of the original remote sensing image;

[0009] The classification network model employs a multi-scale contour wave network based on the Shannon block attention mechanism.

[0010] In one embodiment of the present invention, step 1 includes:

[0011] The original remote sensing image is subjected to global contrast normalization processing to obtain the remote sensing image to be tested.

[0012] In one embodiment of the present invention, the multi-scale contourlet network based on the Shannon block attention mechanism includes: an LMS module, a Fovea module, a contourlet filter module, multiple Shannon block attention mechanism modules, a first feature fusion module, multiple second feature fusion modules, and a classification module; wherein...

[0013] The remote sensing image to be tested is input into the classification network model, and the LMS module extracts the LMS color space features of the remote sensing image to be tested.

[0014] The Fovea module extracts multi-scale features from the remote sensing image under test.

[0015] The contourlet filter module performs multi-level decomposition on the remote sensing image under test to obtain contourlet transform features at different scales and in different directions. The Shannon block attention mechanism module enhances the contourlet transform features at the corresponding scale to obtain the corresponding Shannon block attention features.

[0016] The second feature fusion module is used to fuse the fusion features of the previous level with the Shannon block attention features of the current level to obtain the fusion features of the current level; wherein, the fusion features of the first level are obtained by fusing the LMS color space features, the multi-scale features and the Shannon block attention features of the first level through the first feature fusion module; the fusion features of the last level are used as the input of the classification module, and the classification module performs classification prediction on the input fusion features to obtain the scene classification result.

[0017] In one embodiment of the present invention, the LMS module extracts the LMS color space features of the remote sensing image to be tested, including:

[0018] The remote sensing image to be tested is converted from RGB space to CIE XYZ space;

[0019] The remote sensing image to be tested is converted from CIE XYZ space to LMS space to obtain the LMS color space features of the remote sensing image to be tested.

[0020] In one embodiment of the present invention, the Fovea module performs Chevron sampling on the remote sensing image to be tested to obtain multi-scale features of the remote sensing image to be tested.

[0021] In one embodiment of the present invention, the contourlet filter module includes a scale filter and a direction filter. The contourlet filter module uses the scale filter and the direction filter to iteratively convolve with the input image to decompose and obtain contourlet transform features of different directions at different scales.

[0022] The scaling filter is used to decompose the input image to obtain corresponding low-pass and high-pass components. The high-pass component is input to the directional filter for directional subband decomposition, and the low-pass component is input to the next-level scaling filter for decomposition.

[0023] In one embodiment of the present invention, the Shannon block attention mechanism module includes a channel attention module and a spatial attention module, wherein,

[0024] The channel attention module includes three branches: average pooling, max pooling, and Shannon pooling, which are used to extract the channel attention vector of the input contourlet transform features.

[0025] The channel attention vector is multiplied by the contour wave transform feature to obtain the channel thinning feature. The spatial attention module extracts features from the input channel thinning feature to obtain the Shannon block attention feature.

[0026] In one embodiment of the present invention, the first feature fusion module is a first cascaded layer;

[0027] The second feature fusion module includes a first convolutional layer, a second convolutional layer, a contourlet pooling layer, and a second cascaded layer, wherein,

[0028] The previous level's fused features are passed through the first convolutional layer and the contourlet pooling layer to obtain the contourlet pooling features; the second cascaded layer fuses the contourlet pooling features and the Shannon block attention features of the current level that have passed through the corresponding second convolutional layer to obtain the current level's fused features.

[0029] In one embodiment of the present invention, the classification module includes a cascaded third convolutional layer, a global average pooling layer, a fully connected layer, and a softmax function layer.

[0030] In one embodiment of the present invention, the training process of the classification network model includes:

[0031] Step a: Acquire multiple remote sensing images including various categories, assign a corresponding real category label to each remote sensing image, and perform standardization preprocessing on the remote sensing images;

[0032] Step b: Use the standardized preprocessed remote sensing images and their corresponding ground truth class labels as training samples to form a training dataset;

[0033] Step c: Construct a multi-scale contour wave network based on the Shannon block attention mechanism;

[0034] Step d: Train the multi-scale contour wave network based on the Shannon block attention mechanism according to the training dataset until a preset cutoff condition is reached to obtain a trained classification network model; wherein, during the training process, the network parameters are learned by using the backpropagation algorithm of the error between the true class label of the training sample and the predicted class label output by the corresponding multi-scale contour wave network based on the Shannon block attention mechanism.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] The image classification method based on the Shannon block attention mechanism of this invention utilizes a multi-scale contourlet network to classify remote sensing images. This network leverages the advantages of convolutional neural networks in the spatial domain and Contourlet feature learning in the frequency domain to extract features from remote sensing images at multiple scales and in multiple directions, thereby improving classification accuracy. Furthermore, Contourlet pooling effectively preserves edge and texture details of the image, which also enhances the classification effect of remote sensing images to some extent.

[0037] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0038] Figure 1 This is a flowchart of an image classification method based on a Shannon block attention mechanism multi-scale contour wave network provided by an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of a multi-scale contour wave network based on the Shannon block attention mechanism provided in an embodiment of the present invention;

[0040] Figure 3 This is an exploded schematic diagram of a contourlet filter module provided in an embodiment of the present invention. Detailed Implementation

[0041] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail, with reference to the accompanying drawings and specific embodiments, an image classification method based on a multi-scale contour wave network of Shannon block attention mechanism proposed in accordance with the present invention.

[0042] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0043] Please see Figure 1 , Figure 1 This is a flowchart of an image classification method based on a Shannon block attention mechanism using a multi-scale contourlet network, as provided in this embodiment of the invention. The image classification method based on a Shannon block attention mechanism using a multi-scale contourlet network includes:

[0044] Step 1: Acquire the original remote sensing image, perform standardization preprocessing on the original remote sensing image to obtain the remote sensing image to be tested;

[0045] In an optional implementation, step 1 includes: performing global contrast normalization processing on the original remote sensing image to obtain the remote sensing image to be tested.

[0046] It should be noted that in other implementations, other normalization methods can also be used to standardize and preprocess the original remote sensing images, and no restrictions are imposed here.

[0047] Step 2: Input the remote sensing image to be tested into the trained classification network model to obtain the scene classification result of the original remote sensing image;

[0048] In one alternative implementation, the classification network model employs a multi-scale contour wave network based on the Shannon block attention mechanism.

[0049] Please refer to the above. Figure 2 The schematic diagram shown in the figure illustrates a multi-scale contour wave network based on the Shannon block attention mechanism provided by an embodiment of the present invention. As shown in the figure, the multi-scale contour wave network based on the Shannon block attention mechanism includes: an LMS module, a Fovea module, a contourlet filter module, multiple Shannon block attention mechanism modules, a first feature fusion module, multiple second feature fusion modules, and a classification module.

[0050] Furthermore, the processing procedure of the remote sensing image to be tested in the classification network model is explained. After the remote sensing image to be tested is input into the classification network model, the LMS module extracts the LMS color space features of the remote sensing image to be tested.

[0051] In this embodiment, the remote sensing image to be measured is first converted from RGB space to CIE XYZ space; optionally, the conversion formula is as follows:

[0052]

[0053] The remote sensing image to be tested is then converted from the CIE XYZ space to the LMS space to obtain the LMS color space features of the remote sensing image. Optionally, the conversion formula is as follows:

[0054]

[0055] It should be noted that R, G, and B in formula (1) need to be normalized and gamma-corrected first.

[0056] In one optional implementation, the Fovea module extracts multi-scale features from the remote sensing image to be tested. Optionally, the Fovea module performs Chevron sampling on the remote sensing image to be tested to obtain multi-scale features of the remote sensing image to be tested.

[0057] For example, suppose the output O of the remote sensing image to be measured is cropped exponentially in the horizontal and vertical directions to a multi-scale manner. i (i = 1, 2, ..., s), then the multi-scale features generated by Chevron sampling are:

[0058]

[0059] In the formula, s represents the number of scales, and c represents the Cevron parameter. Represents region O i Cut off O i-c .

[0060] In one optional implementation, the contourlet filter module performs multi-level decomposition on the remote sensing image to be measured to obtain contourlet transform features at different scales and in different directions, and the Shannon block attention mechanism module enhances the contourlet transform features at the corresponding scale to obtain the corresponding Shannon block attention features.

[0061] The essence of a contourlet filter module is the process of iteratively convolving a signal with a Laplacian pyramid (LP) and directional filter banks (DFB).

[0062] In this embodiment, the contourlet filter module includes a scaling filter ( Figure 2 The pyramid filter and directional filter are used in the contourlet filter module to iteratively convolve with the input image using the scale filter and directional filter, decomposing it to obtain the contour wave transform features at different scales and in different directions.

[0063] Please see as follows Figure 3 The diagram shown illustrates the decomposition of a contourlet filter module according to an embodiment of the present invention. As shown, the scaling filter is used to decompose the input image into corresponding low-pass and high-pass components. The high-pass component is input to the directional filter for directional subband decomposition, and the low-pass component is input to the next-stage scaling filter for further decomposition.

[0064] For example, let the input of the scaling filter be X, and let X... l,i and X h,i Let X represent the low-pass and high-pass components of the output of the i-th scale filter, respectively. l,0 =X. Therefore, the output of the (i+1)th level contourlet filter can be written as:

[0065]

[0066] Among them, F LP For maxflat Laplace decomposition, F DFB This is a dmaxflat7 directional filter. p is the downsampling factor, and the subscript h_bds represents the bandpass direction subband in the spectral domain.

[0067] In this embodiment, L = 4 scale and direction filters are used, nlevels = [0, 3, 3, 3], where 0 represents two-dimensional wavelet decomposition, and l = 1, ..., L.

[0068] It should be noted that in this embodiment, the size of the Contourlet decomposition feature map needs to be adjusted according to the actual situation to facilitate concatenation with the corresponding features of the network. In other embodiments, other scales, Laplace decompositions, and directional filters may be used, and no restrictions are imposed here.

[0069] Optionally, the contourlet filter extracts the contourlet transform features corresponding to the input image using a coefficient matrix. The coefficient matrix represents... The size was adjusted to M 2 ,in

[0070] In one optional implementation, the Shannon block attention mechanism module includes a channel attention module ( Figure 2 CAM) and spatial attention module ( Figure 2 The SAM (Shannon Block Attention) module includes three branches: average pooling, max pooling, and Shannon pooling, which are used to extract the channel attention vector of the input contourlet transform features. The channel attention vector is multiplied by the contourlet transform features to obtain the channel thinning features. The spatial attention module extracts features from the input channel thinning features to obtain the Shannon block attention features.

[0071] For example, let F be the contourlet transform feature input to the Shannon block attention mechanism module, and M be the channel attention vector extracted by the channel attention module. c (F) can be represented as:

[0072]

[0073] Where σ represents the sigmoid activation function, and MLP is a multilayer perceptron with one hidden layer. and These are the weights of the perceptron. For average pooling characteristics, For max pooling features, This is characteristic of Shannon pondification.

[0074] Obtained through Shannon pooling, and represented as:

[0075]

[0076] in, E represents the Shannon entropy of the k-th directional subband coefficients of the contour wave decomposition at a specific scale l. l p represents the Shannon entropy of all directional subband coefficients of the contour wave decomposition at a specific scale l. k This represents the probability corresponding to the coefficient of the k-th direction sub-band.

[0077] Channel attention vector M c (F) is multiplied by the contour wave transform feature F to obtain the channel refinement feature F′, and the Shannon block attention feature M is extracted by the spatial attention module. s (F′) can be represented as:

[0078]

[0079] in, This represents the average pooling feature along the channel dimension. denoted as max pooling feature along the channel dimension, and f represents the convolution operation.

[0080] In an optional implementation, the second feature fusion module is used to fuse the fusion features of the previous level with the Shannon block attention features of the current level to obtain the fusion features of the current level; the fusion features of the first level are obtained by fusing the LMS color space features, multi-scale features and the Shannon block attention features of the first level through the first feature fusion module.

[0081] In this embodiment, the first feature fusion module is a first cascaded layer; the second feature fusion module includes a first convolutional layer ( Figure 2(middle conv1 / conv2), second convolutional layer ( Figure 2 conv 1-a / conv 2-a ), contourlet pooling layer ( Figure 2 The first convolutional layer (CP1 / CP2) and the second concatenated layer are used to obtain the contour pooling feature by passing the first convolutional layer and the contourlet pooling layer in sequence. The second concatenated layer fuses the contour pooling feature and the Shannon block attention feature of the current level after passing through the corresponding second convolutional layer to obtain the fused feature of the current level.

[0082] Contourlet pooling layers, in particular, preserve as much detail as possible to obtain invariant structural features. Let the input features of the contourlet pooling layer be... Perform contour wave decomposition on it T contourlet (·), to obtain the corresponding low-pass subband and Qualcomm sub-band The pooling process can be represented as:

[0083]

[0084] In an optional implementation, the fusion features of the last level are used as input to the classification module, which performs classification prediction on the input fusion features to obtain the scene classification result.

[0085] Optionally, the classification module includes a cascaded third convolutional layer ( Figure 2 conv3), global average pooling layer (not shown in the figure), fully connected layer ( Figure 2 The layers are FC1 / FC2 / FC3 and softmax function layers (not shown in the figure).

[0086] In this embodiment, the fusion feature of the last stage is denoted as F. concat , will F concat Further feature extraction is performed to obtain the final feature representation. For example, a 3×3 convolution operation is first performed. 3×3 (·), and then feed it sequentially into a 7×7 global average pooling layer and a fully connected layer. The corresponding operation can be represented as:

[0087]

[0088] Where Y represents the output of the fully connected layer.

[0089] Based on the input Y of the fully connected layer, the softmax function is used to perform classification prediction, and the prediction result (i.e., the scene classification result of the original remote sensing image) is as follows:

[0090]

[0091] Where K represents the number of categories, y j This indicates the predicted category label.

[0092] In this embodiment, before classifying the remote sensing image to be measured using the classification network model, it needs to be trained. The training process of the classification network model in this embodiment is further explained, specifically including the following steps:

[0093] Step a: Acquire multiple remote sensing images including various categories, assign a corresponding real category label to each remote sensing image, and perform standardization preprocessing on the remote sensing images;

[0094] Optionally, each remote sensing image may be subjected to standardized preprocessing using methods including but not limited to global contrast normalization.

[0095] Step b: Use the standardized preprocessed remote sensing images and their corresponding ground truth class labels as training samples to form a training dataset;

[0096] Step c: Construct a multi-scale contour wave network based on the Shannon block attention mechanism;

[0097] In this embodiment, the parameters of the corresponding contourlet filter (including the number of levels of the scale filter, the number of decomposition directions of the directional filter, and the filter type) and the number of Shannon block attention mechanism modules connected to the contourlet filter are set according to the set scale value. See [link to specific network structure] for details. Figure 2 As shown.

[0098] Step d: Train the multi-scale contour wave network based on the Shannon block attention mechanism using the training dataset until the preset cutoff condition is met, and obtain the trained classification network model.

[0099] During training, the multi-scale contour wave network based on the Shannon block attention mechanism processes the training samples in a similar way to the classification network model in processing the remote sensing images to be tested, and will not be described in detail here.

[0100] During training, the network parameters are learned by backpropagation of the error between the true class labels of the training samples and the predicted class labels of the corresponding multi-scale contour wave network based on the Shannon block attention mechanism.

[0101] In this embodiment, the preset cutoff condition is reaching a preset number of training iterations.

[0102] The image classification method based on the Shannon block attention mechanism of this invention utilizes a multi-scale contourlet network to classify remote sensing images. This network leverages the advantages of convolutional neural networks in the spatial domain and Contourlet feature learning in the frequency domain to extract features from remote sensing images at multiple scales and in multiple directions, thereby improving classification accuracy. Furthermore, Contourlet pooling effectively preserves edge and texture details of the image, which also enhances the classification effect of remote sensing images to some extent.

[0103] It is worth noting that the multi-scale contour wave network based on the Shannon block attention mechanism proposed in this embodiment is not limited to the classification of remote sensing images. It can also be used to classify handwritten digit images or texture images after training with handwritten datasets (MNIST) and texture datasets (kth-tips2-b or CUReT).

[0104] Furthermore, this embodiment also illustrates the effectiveness of the image classification method based on the Shannon block attention mechanism multi-scale contour wave network through simulation experiments.

[0105] 1. Simulation conditions

[0106] The simulation test platform was built on an HP-Z840 high-performance graphics workstation running Ubuntu 16.04 LTS, equipped with two NVIDIA GeForce GTX 1080 graphics cards and an Intel Xeon E5 processor, with 64GB of dedicated video memory. The computer software was Caffe.

[0107] The data used in the simulation included handwritten datasets (MNIST), texture datasets (kth-tips2-b, CUReT), and remote sensing scene classification datasets (UC Merced (UCM), AID, NWPU-RESISC45).

[0108] 2. Evaluation of classification results

[0109] In the evaluation of image classification simulation experiments, the overall accuracy (OA) is used for assessment, and its calculation formula is as follows:

[0110]

[0111] TP, TN, FN, and FP represent the number of true positive, true negative, false negative, and false positive samples, respectively.

[0112] 3. Simulation Results

[0113] (1) The effectiveness of the image classification method based on the Shannon block attention mechanism multi-scale contour wave network was trained and validated on the classic handwritten image classification dataset MNIST. The CIFAR dataset includes 60,000 training samples and 10,000 test samples, with a total of 10 classes. No pre-training or data augmentation was performed in the experiment.

[0114] The classification results of this invention compared with other methods are shown in Table 1. As can be seen from Table 1, this invention has a higher classification accuracy.

[0115] Table 1. Classification results of this invention and other comparative methods on the MNIST dataset.

[0116]

[0117]

[0118] (2) The effectiveness of the image classification method based on the Shannon block attention mechanism multi-scale contour wave network was trained and validated on the classic texture image classification datasets kth-tips2-b and CUReT. The kth-tips2-b dataset includes four training sample sets, one of which was randomly selected as the training set, and the remaining three as the test set. There are a total of 11 classes. The CUReT dataset includes 92 images per class. 46 images were used for training, and the rest were used for testing, for a total of 61 classes. No pre-training or data augmentation was performed in the experiments.

[0119] The classification results of this invention compared with other methods are shown in Tables 2 and 3. As can be seen from Tables 1 and 2, the classification accuracy of this invention is better than that of other comparison methods.

[0120] Table 2. Classification results of this invention and other comparative methods on the kth-tips2-b dataset.

[0121] method Accuracy (%) AlexNet 48.3 T-CNN 48.7 Wavelet CNN 63.7 C-CNN 70.31 This invention 70.54

[0122] Table 3. Classification results of this invention and other comparative methods on the CURET dataset.

[0123] method Accuracy (%) CISDD 99.3 AlexNet 99.4 T-CNN 99.5 WMACapsNet 99.6 This invention 99.7

[0124] (3) The effectiveness of the image classification method based on the Shannon block attention mechanism multi-scale contour wave network was trained and validated on the classic remote sensing scene classification datasets UC Merced, AID, and RSSCN7. The UC Merced dataset contains 100 images per class. 80 images were used for training, and the remaining 20 were used for testing, for a total of 21 classes. For the AID dataset, 50% of the samples from each class were randomly selected for training, and the remainder were used for testing, for a total of 30 classes. For the RSSCN7 dataset, 50% of the samples from each class were randomly selected for training, and the remainder were used for testing, for a total of 7 classes. No pre-training or data augmentation was performed in the experiments.

[0125] The classification results of this invention compared with other methods are shown in Table 4. As can be seen from Table 4, this invention has better feature representation ability on remote sensing scene datasets UCM, AID and RSCN7, thereby effectively improving the classification accuracy of remote sensing scenes.

[0126] Table 4. Classification results (%) of this invention and other comparative methods on the UC Merced, AID, and RSCN7 datasets.

[0127] method UCM AID RSSCN7 CIPAL 96.58 91.22 - RSFJR 97.21 96.18 - TEX-Nets-LF 97.72 95.73 94.0 SE-MDPMNet 98.95 97.14 94.71 This invention 99.04 97.44 95.05

[0128] (4) The effectiveness of the image classification method based on the Shannon block attention mechanism multi-scale contour wave network was trained and validated on the classic remote sensing scene classification dataset NWPU-RESISC45. In the NWPU-RESISC45 dataset, 10% and 20% of the samples of each class were randomly selected for training, and the remaining 90% and 80% were used for testing, for a total of 45 classes. No pre-training or data augmentation was performed in the experiment.

[0129] The classification results of this invention compared with other methods are shown in Table 5. As can be seen from Table 5, this invention can capture discriminative remote sensing scene features and achieve good classification accuracy on training sets of different proportions.

[0130] Table 5. Results (%) of this invention and other comparative methods on the NWPU-RESISC45 dataset.

[0131]

[0132]

[0133] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0134] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An image classification method based on a multi-scale contour wave network with a Shannon block attention mechanism, characterized in that, include: Step 1: Acquire the original remote sensing image, and perform standardization preprocessing on the original remote sensing image to obtain the remote sensing image to be tested; Step 2: Input the remote sensing image to be tested into the trained classification network model to obtain the scene classification result of the original remote sensing image; The classification network model employs a multi-scale contourlet network based on the Shannon block attention mechanism. This multi-scale contourlet network includes: an LMS module, a Fovea module, a contourlet filter module, multiple Shannon block attention mechanism modules, a first feature fusion module, multiple second feature fusion modules, and a classification module. The remote sensing image to be tested is input into the classification network model, and the LMS module extracts the LMS color space features of the remote sensing image to be tested. The Fovea module extracts multi-scale features from the remote sensing image under test. The contourlet filter module performs multi-level decomposition on the remote sensing image under test to obtain contourlet transform features at different scales and in different directions. The Shannon block attention mechanism module enhances the contourlet transform features at the corresponding scale to obtain the corresponding Shannon block attention features. The second feature fusion module is used to fuse the fusion features of the previous level with the Shannon block attention features of the current level to obtain the fusion features of the current level; wherein, the fusion features of the first level are obtained by fusing the LMS color space features, the multi-scale features and the Shannon block attention features of the first level through the first feature fusion module; the fusion features of the last level are used as the input of the classification module, and the classification module performs classification prediction on the input fusion features to obtain the scene classification result.

2. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, Step 1 includes: The original remote sensing image is subjected to global contrast normalization processing to obtain the remote sensing image to be tested.

3. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The LMS module extracts the LMS color space features of the remote sensing image under test, including: The remote sensing image to be tested is converted from RGB space to CIE XYZ space; The remote sensing image to be tested is converted from CIE XYZ space to LMS space to obtain the LMS color space features of the remote sensing image to be tested.

4. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The Fovea module performs Chevron sampling on the remote sensing image to be tested, thereby obtaining the multi-scale features of the remote sensing image to be tested.

5. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The contourlet filter module includes a scale filter and a direction filter. The contourlet filter module uses the scale filter and the direction filter to iteratively convolve with the input image to decompose and obtain contourlet transform features at different scales and in different directions. The scaling filter is used to decompose the input image to obtain corresponding low-pass and high-pass components. The high-pass component is input to the directional filter for directional subband decomposition, and the low-pass component is input to the next-level scaling filter for decomposition.

6. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The Shannon block attention mechanism module includes a channel attention module and a spatial attention module, wherein... The channel attention module includes three branches: average pooling, max pooling, and Shannon pooling, which are used to extract the channel attention vector of the input contourlet transform features. The channel attention vector is multiplied by the contour wave transform feature to obtain the channel thinning feature. The spatial attention module extracts features from the input channel thinning feature to obtain the Shannon block attention feature.

7. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The first feature fusion module is the first cascaded layer; The second feature fusion module includes a first convolutional layer, a second convolutional layer, a contourlet pooling layer, and a second cascaded layer, wherein, The previous level's fused features are passed through the first convolutional layer and the contourlet pooling layer to obtain the contourlet pooling features; the second cascaded layer fuses the contourlet pooling features and the Shannon block attention features of the current level that have passed through the corresponding second convolutional layer to obtain the current level's fused features.

8. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The classification module includes a cascaded third convolutional layer, a global average pooling layer, a fully connected layer, and a softmax function layer.

9. The image classification method based on Shannon block attention mechanism multi-scale contour wave network according to claim 1, characterized in that, The training process of the classification network model includes: Step a: Acquire multiple remote sensing images including various categories, assign a corresponding real category label to each remote sensing image, and perform standardization preprocessing on the remote sensing images; Step b: Use the standardized preprocessed remote sensing images and their corresponding ground truth class labels as training samples to form a training dataset; Step c: Construct a multi-scale contour wave network based on the Shannon block attention mechanism; Step d: Train the multi-scale contour wave network based on the Shannon block attention mechanism according to the training dataset until the preset cutoff condition is reached to obtain the trained classification network model; wherein, during the training process, the network parameters are learned by using the backpropagation algorithm of the error between the true class label of the training sample and the predicted class label output by the corresponding multi-scale contour wave network based on the Shannon block attention mechanism.

Images