Image processing method, apparatus, system, device and readable storage medium

Abstract

This application relates to an image processing method, apparatus, system, device, and readable storage medium. The method is applied to an image processing system including a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder. The multi-layer cascaded encoder and decoder are connected in a skip connection. The method includes: acquiring an initial synthetic aperture radar (SAR) image; performing layer-by-layer feature extraction on the initial SAR image by the multi-layer cascaded encoder to obtain multi-level image features; for each level of image features, performing multi-scale frequency domain enhancement processing based on the image features of that level by the corresponding wavelet decomposition module to obtain the enhanced features corresponding to that level; wherein the wavelet basis included in the wavelet decomposition module is adapted to the image features; and performing layer-by-layer decoding on the enhanced features corresponding to each level to obtain the image processing result. This method can improve the accuracy of image processing.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image processing method, apparatus, system, device, and readable storage medium. Background Technology

[0002] With the development of Synthetic Aperture Radar (SAR) remote sensing technology, high-resolution SAR images have been widely used in disaster monitoring, resource exploration, and military reconnaissance. The accompanying SAR image segmentation technology has also evolved accordingly. SAR image segmentation can extract the contour and category information of ground targets from complex radar echo data, providing a foundation for subsequent target interpretation and decision-making. However, unlike optical images, SAR images rely on the backscattering coefficients of ground objects for imaging, naturally exhibiting inherent defects such as severe speckle noise, complex image textures, and difficulty in representing small targets. This places higher demands on the robustness and accuracy of segmentation algorithms.

[0003] In related technologies, segmentation of SAR images mainly relies on statistical models and handcrafted features, as well as segmentation methods based on convolutional neural networks. However, these methods suffer from severe speckle noise interference and blurred edges of small targets, resulting in poor processing accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide an image processing method, apparatus, system, device, and readable storage medium that can improve the accuracy of image processing in response to the above-mentioned technical problems.

[0005] In a first aspect, this application provides an image processing method applied to an image processing system, the image processing system including a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder, wherein the multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection, the method comprising:

[0006] An initial synthetic aperture radar image is acquired, and the multi-layer cascaded encoder performs layer-by-layer feature extraction on the initial synthetic aperture radar image to obtain multi-level image features;

[0007] For each level of image features, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level; wherein, the wavelet basis included in the wavelet decomposition module is adapted to the image features;

[0008] The multi-layered decoder decodes the enhanced features corresponding to each layer layer by layer to obtain the image processing result.

[0009] Secondly, this application also provides an image processing apparatus applied to an image processing system, the image processing system including a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder, wherein the multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection, including:

[0010] The first processing module is used to acquire an initial synthetic aperture radar image, and the multi-layer cascaded encoder performs layer-by-layer feature extraction on the initial synthetic aperture radar image to obtain multi-level image features.

[0011] The second processing module is used to perform multi-scale frequency domain enhancement processing and feature fusion based on the image features of each level by the wavelet decomposition module corresponding to that level, to obtain the fused features corresponding to that level; wherein, the wavelet basis included in the wavelet decomposition module is determined based on the image features;

[0012] The third processing module is used to perform layer-by-layer decoding and feature enhancement on the fusion features corresponding to each of the multi-layer cascaded decoders to obtain the image processing result.

[0013] Thirdly, this application also provides an image processing system, which includes a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder; the multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection, wherein:

[0014] The multi-layer cascaded encoder is used to extract features layer by layer from the initial synthetic aperture radar image to obtain multi-level image features;

[0015] The wavelet decomposition module is used to perform multi-scale frequency domain enhancement processing and feature fusion based on the image features of each level, thereby obtaining the fused features corresponding to that level; wherein, the wavelet basis included in the wavelet decomposition module is determined based on the image features;

[0016] The multi-layer cascaded decoder is used to perform layer-by-layer decoding and feature enhancement on the fusion features corresponding to each layer to obtain the image processing result.

[0017] Fourthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0018] Fifthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0019] Sixthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0020] The aforementioned image processing method, apparatus, system, device, and readable storage medium employ a multi-layer cascaded encoder to extract features from radar images layer by layer. This multi-layered encoder captures multi-scale, multi-level image features, ranging from details to the overall picture. Subsequently, a wavelet decomposition module is used to perform multi-scale frequency domain enhancement on the image features at each level, thereby suppressing noise and enhancing the capture of high-frequency texture information in the frequency domain. Simultaneously, during the multi-scale frequency domain enhancement process in the wavelet decomposition module, since the wavelet basis is adapted to the image features, it can be dynamically adjusted according to different levels and content of image features, achieving more accurate frequency domain decomposition and enhancement. This avoids the problems of poor adaptability and limited enhancement effect caused by a fixed wavelet basis. Finally, a multi-layer cascaded decoder decodes the enhanced features corresponding to each level layer by layer to obtain the image processing result, achieving coarse-to-fine image reconstruction. Furthermore, since the encoder and decoder use a skip connection, the original image features of each level of the encoder can be directly reused during the decoding stage, effectively preserving image details, reducing information loss, and further improving image processing accuracy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an application environment diagram of an image processing method in one embodiment;

[0023] Figure 2 This is a flowchart illustrating an image processing method in one embodiment;

[0024] Figure 3 This is a schematic diagram of the system structure of an image processing system in one embodiment;

[0025] Figure 4 This is a schematic diagram illustrating the construction process of the wavelet basis decomposition module in one embodiment;

[0026] Figure 5 This is a schematic diagram of the image processing system in yet another embodiment;

[0027] Figure 6 This is a schematic diagram of the data processing flow of the wavelet basis decomposition module in one embodiment;

[0028] Figure 7 This is a visual schematic diagram of different frequency band components involved in one embodiment;

[0029] Figure 8 This is a schematic diagram illustrating the effect of frequency domain enhancement processing in one embodiment.

[0030] Figure 9 This is a schematic diagram of the structure of a geometry-aware local-global attention module involved in one embodiment;

[0031] Figure 10 This is a structural block diagram of an image processing device in one embodiment;

[0032] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0035] The image processing method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be integrated onto server 102 or placed on a cloud or other network server. Server 102 can also deploy an image processing system, which includes a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder, with skip connections between the multi-layer cascaded encoder and the multi-layer cascaded decoder.

[0036] The user can input an initial synthetic aperture radar (SAR) image to be processed through terminal 101, generate an image processing task for the initial SAR image, and send an image processing request to server 102. After receiving the image processing request, server 102 obtains the initial SAR image by parsing the image processing request, and performs layer-by-layer feature extraction on the initial SAR image by a multi-layer cascaded encoder to obtain multi-level image features. For each level of image features, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level. The wavelet basis included in the wavelet decomposition module is adapted to the image features. The multi-layer cascaded decoder decodes the enhanced features corresponding to each level layer by layer to obtain the image processing result. Finally, server 102 can return the image processing result to terminal 101, and terminal 101 can display the image processing result to the user.

[0037] Terminal 101 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, and projection equipment. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses.

[0038] Server 102 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0039] In one exemplary embodiment, such as Figure 2 As shown, an image processing method is provided, which is applied to... Figure 1 Taking a server as an example, an image processing system can be deployed on the server. This system includes a multi-layered cascaded encoder, a wavelet decomposition module, and a multi-layered cascaded decoder. The multi-layered cascaded encoder and decoder are interconnected. The image processing method includes steps 201 to 203. Wherein:

[0040] Step 201: Obtain the initial synthetic aperture radar image, and perform layer-by-layer feature extraction on the initial synthetic aperture radar image by a multi-layer cascaded encoder to obtain multi-level image features.

[0041] In some embodiments, the initial synthetic aperture radar image may be uploaded by the terminal device; or the initial synthetic aperture radar image may be actively retrieved by the server from the acquisition device. The specific acquisition method is not limited here.

[0042] Among them, a multi-layer cascaded encoder refers to a feature extraction structure consisting of multiple sequentially connected and progressively advanced coding network layers, with each coding layer connected in a cascade manner.

[0043] Furthermore, the extraction process of the multi-layer cascaded encoder includes: inputting the initial synthetic aperture radar image into the multi-layer cascaded encoder, performing feature extraction on the image through each coding layer in sequence, such as convolution, downsampling, nonlinear transformation, etc., and using the feature map output by the previous layer as the input of the next layer, thereby gradually extracting the middle-level semantic features and high-level abstract features to obtain image feature maps of different scales and levels.

[0044] Multi-level image features refer to the image features output by each layer of an encoder in a multi-level cascaded encoder.

[0045] Step 202: For each level of image features, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level; wherein, the wavelet basis contained in the wavelet decomposition module is adapted to the image features.

[0046] In this system, the encoders in the multi-layer cascaded array are connected to the wavelet decomposition module. Alternatively, each level of the encoder in the multi-layer cascaded array can be connected to the wavelet decomposition module separately. Or, some levels of the encoders in the multi-layer cascaded array can be connected to the wavelet decomposition module. The specific connection method can be set according to actual needs.

[0047] Wavelet decomposition refers to using wavelet transform to decompose the feature map into sub-band components of different frequencies and scales in the spatial domain, which correspond to the low-frequency contour information and high-frequency detail information in the feature, respectively, thereby realizing the hierarchical representation and separation processing of the feature in the frequency domain.

[0048] Wavelet basis refers to the basis functions used in wavelet transform to decompose and reconstruct feature maps. Different wavelet basis has different time-domain and frequency-domain characteristics. Wavelet basis is selected and adapted according to the texture, edge, scale and distribution characteristics of image features to improve the decomposition and enhancement effect.

[0049] Frequency domain enhancement processing refers to the process of converting image features from the spatial domain to the frequency domain, performing filtering, weighting, and gain adjustment on different frequency components such as high-frequency details and low-frequency contours to enhance effective information, suppress noise and redundant components, and then converting them back to the spatial domain, thereby improving feature clarity and expressive effect.

[0050] Multi-scale frequency domain enhancement processing refers to using multiple wavelet bases of different scales to perform wavelet decomposition on image features, and performing enhancement operations such as gain adjustment and filtering optimization at each scale.

[0051] In some embodiments, the wavelet basis is adapted to image features. This can mean that the polarity distribution of the wavelet basis matches the image features, or that the response intensity of the wavelet basis matches the image features; or that both the polarity distribution and the response loudness of the wavelet basis are adapted to the image features. The polarity distribution of the wavelet basis is used to characterize the distribution pattern of positive excitation and negative inhibition generated by the wavelet basis for signals in different directions and regions of the feature map, that is, the sensitive and suppressed regions of the wavelet basis for local features. The response intensity of the wavelet basis is used to characterize the degree of matching and extraction capability of the wavelet basis for corresponding texture, edge, contour and other information in the feature map.

[0052] Clearly, by using a wavelet basis adapted to the image features and performing wavelet transform on the image features, frequency band features can be extracted more accurately, thus obtaining more accurate enhanced features.

[0053] In some embodiments, correction coefficients can be determined by image features, and then the wavelet basis can be adjusted using the correction coefficients to make the adjusted wavelet basis fit the image features.

[0054] In some embodiments, image feature libraries of different levels and matching wavelet basis libraries can be pre-constructed. During feature enhancement, based on the image features of the current level, the optimal matching wavelet basis is adaptively selected from the wavelet basis library, and the selected wavelet basis is used to perform multi-scale frequency domain enhancement processing on the image features.

[0055] In other embodiments, the correction coefficients can also be used as learnable parameters. During model training, the learnable parameters are adaptively updated using the backpropagation algorithm to obtain correction coefficients that are adapted to image features at different levels.

[0056] In some embodiments, considering that wavelet decomposition is mainly used to extract local detail information such as texture and edges, and that as the coding level deepens, the extracted image features are highly abstracted and no longer contain rich local texture and edge information, there is no need to perform wavelet decomposition. That is, in a multi-layer cascaded encoder, it is not necessary for each level to be connected to the wavelet decomposition module. Therefore, in some embodiments, step 202 may include: for the image features of each level, if the encoder of that level is connected to a wavelet decomposition module, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level; if the encoder of that level is not connected to a wavelet decomposition module, the image features of that level are used as the enhanced features corresponding to that level, wherein the wavelet basis contained in the wavelet decomposition module is adapted to the image features;

[0057] Step 203: The multi-layer cascaded decoder decodes the enhanced features corresponding to each layer layer by layer to obtain the image processing result.

[0058] In this context, a multi-layer cascaded decoder corresponds to a multi-layer cascaded encoder, meaning that the encoder and decoder have the same number of layers; a skip connection between a multi-layer cascaded encoder and a multi-layer cascaded decoder refers to a skip connection between encoders and decoders at corresponding layers.

[0059] For example, the encoder includes n layers, where encoding is performed sequentially from the first layer to the nth layer. The first layer is used to perform feature encoding on the initial synthetic aperture radar image, and the output of the first layer is used as the input of the second layer, and so on. At the same time, the decoder includes n layers, where decoding is performed sequentially from the first layer to the nth layer. The nth layer is used to output the final image processing result, and the output of the first layer is used as the input of the second layer. In this case, the encoder of the i-th layer is skipped to the decoder of the (n+1-i)-th layer.

[0060] In some embodiments, a multi-layer cascaded decoder decodes the enhancement features corresponding to each layer layer by layer to obtain the image processing result. The specific process includes: the input of the first layer decoder is the enhancement feature and image feature corresponding to that layer, and the output is the decoded feature of the first layer; the input of the second layer decoder is the enhancement feature and image feature corresponding to that layer as well as the decoded feature of the first layer, and the output is the decoded feature of the second layer, and so on, until the decoded feature output by the last layer decoder is obtained. Based on the decoded feature output by the last layer decoder, the image processing result is obtained.

[0061] For easier understanding, please refer to Figure 3 , Figure 3 A schematic diagram of the system structure of the image processing system according to an embodiment of this application is shown.

[0062] The image processing system includes a four-layer cascaded encoder and a four-layer cascaded decoder, with skip connections between the encoders and decoders at corresponding layers. Figure 3 (The connection relationship is not shown in the image). Specifically, the input to the first-layer encoder is the initial synthetic aperture radar image, and the output of the first-layer encoder is used as the input to the second-layer encoder, and so on. At the same time, the encoders and decoders of the corresponding layers are connected through wavelet basis decomposition modules. The input to the decoder is the output of the previous layer decoder and the output of the corresponding wavelet basis decomposition module. The image processing result is obtained through multi-layer decoding.

[0063] In the aforementioned image processing method, a multi-layer cascaded encoder is used to extract features from radar images layer by layer. The multi-layered encoder captures multi-scale, multi-level image features from details to the global picture. Then, a wavelet decomposition module is used to perform multi-scale frequency domain enhancement on the image features at each level, thereby suppressing noise and enhancing the capture of high-frequency texture information in the frequency domain. Simultaneously, during the multi-scale frequency domain enhancement process in the wavelet decomposition module, since the wavelet basis is adapted to the image features, it can be dynamically adjusted according to different levels and content of image features, achieving more accurate frequency domain decomposition and enhancement. This avoids the problems of poor adaptability and limited enhancement effect caused by a fixed wavelet basis. Finally, a multi-layer cascaded decoder decodes the enhanced features corresponding to each level layer by layer to obtain the image processing result, achieving coarse-to-fine image reconstruction. Furthermore, since the encoder and decoder use a skip connection, the original image features of each level of the encoder can be directly reused in the decoding stage, effectively preserving image details, reducing information loss, and further improving image processing accuracy.

[0064] In one exemplary embodiment, such as Figure 4 As shown, Figure 4 The construction process of the wavelet basis decomposition module is illustrated, which includes steps 301 to 303. Wherein:

[0065] Step 301: Construct an initial wavelet basis; wherein the initial wavelet basis includes multiple different frequency band components.

[0066] In some embodiments, the plurality of different frequency band components include at least two of the following: low-low approximation component (Low-Low, LL), horizontal edge component (Low-High, LH), vertical edge component (High-Low, HL), and diagonal edge component (High-High, HH).

[0067] For example, the initial wavelet basis may include a set of basic 2×2 Haar wavelet kernels, each corresponding to a different frequency band component.

[0068] Step 302: Perform interpolation based on the initial wavelet basis to obtain multiple wavelet bases at different scales.

[0069] Interpolation refers to calculating and supplementing the parameter values ​​at intermediate positions between known parameter points of the initial wavelet basis using an interpolation algorithm, thereby expanding the convolution kernel of the initial wavelet basis from its original size to the target size, such as from 2×2 to 4×4. Interpolation can maintain the continuous and smooth spatial distribution characteristics of the expanded wavelet basis, and keep it consistent with the polarity distribution and response intensity of the initial wavelet basis.

[0070] In some embodiments, bilinear interpolation can be used to obtain multiple wavelet bases at different scales; for example, the initial wavelet base corresponds to a 2×2 wavelet kernel, and through interpolation, wavelet kernels at multiple scales such as 4×4 and 8×8 can be obtained, which are multiple wavelet bases at different scales.

[0071] Step 303: Add correction parameters to each wavelet basis to obtain the wavelet decomposition module; wherein, the correction parameters include at least one of the position perturbation parameters and the intensity scaling parameters, and the correction parameters are used to adaptively adjust the wavelet basis according to the image features.

[0072] Among them, the position perturbation parameter refers to the parameter used to make a small offset adjustment to the spatial position of the wavelet basis. By changing the response distribution position of the wavelet basis on the feature map, the wavelet basis can be accurately aligned with the edge, texture and other structures in the image features, thereby improving the matching degree between the wavelet basis and local features.

[0073] Intensity scaling parameters are parameters used to weight and adjust the response amplitude of a wavelet basis. By adjusting the output response of the wavelet basis, the feature components can be enhanced or suppressed, thereby optimizing the extraction effect of the wavelet basis on features of different frequency bands and intensities.

[0074] In some embodiments, the correction parameters serve as learnable model parameters, enabling the image processing system to learn the correlation between image features and correction parameters during model training. For example, inputting a sample radar image and a reference image processing result, the image processing system encodes, performs wavelet enhancement, and decodes the sample radar image. The segmentation head then outputs a predicted image processing result. A loss is calculated based on the reference and predicted image processing results, and the loss gradient is backpropagated through this loss, thereby iterating the model parameters in the image processing system. In subsequent applications, the image processing system can dynamically adjust different correction parameters for different input SAR images based on the learned correlation between image features and correction parameters, achieving adaptation between the correction parameters and image features.

[0075] In the above embodiments, multiple wavelet bases of different scales can be generated by interpolation, thereby adapting to different receptive fields and texture granularities. At the same time, by introducing correction parameters for each wavelet base, the spatial distribution and response intensity of the wavelet base can be adaptively adjusted according to the image features, thereby improving the feature enhancement and decomposition effect.

[0076] In some embodiments, for each level of image features, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level, including:

[0077] For each level of image features, parameter inference is performed based on the image features of that level to obtain the target correction parameters corresponding to that level.

[0078] In some embodiments, as mentioned in the foregoing embodiments, the correlation between image features and correction parameters can be acquired adaptively during model training.

[0079] In other embodiments, multiple reference image features and reference correction parameters corresponding to each reference image feature may be pre-established. Then, for each level of image features, the reference image feature closest to the image feature of that level is determined by calculating the feature distance (or feature similarity). The target correction parameter is obtained based on the feature distance (or feature similarity) and the reference correction parameter corresponding to the reference image feature.

[0080] Based on the target correction parameters, the wavelet bases at multiple scales included in the wavelet decomposition module are deformed to obtain the target wavelet bases at multiple scales.

[0081] For example, by substituting the target correction parameters into the wavelet basis functions corresponding to the wavelet basis at multiple scales, the target wavelet basis at multiple scales can be obtained.

[0082] The target wavelet basis at multiple scales is used to perform frequency domain decomposition at different scales for the image features at that level, thereby obtaining frequency domain features at different scales.

[0083] In some embodiments, the frequency domain decomposition at each scale can be performed on multiple different frequency band components at that scale, such as low-frequency approximate components, horizontal edge components, vertical edge components, and diagonal edge components. Then, the frequency domain decomposition results of different frequency band components are weighted to obtain the frequency domain features at that scale.

[0084] Furthermore, the weights corresponding to the frequency domain decomposition results of different frequency band components can be adaptively learned through the backpropagation algorithm of model training.

[0085] By weighting the frequency domain features at different scales, the enhanced features corresponding to each level are obtained.

[0086] For example, the frequency domain features at different scales are weighted and summed to obtain the enhanced features corresponding to the level; the weights corresponding to different scales can be adaptively learned through the backpropagation algorithm of the model training.

[0087] In the above embodiments, different target correction parameters can be determined for different image features, thereby dynamically adjusting the corresponding wavelet decomposition module using the target correction parameters to match the image features and improve the accuracy of wavelet decomposition. At the same time, wavelet bases of multiple scales can correspond to different receptive fields and texture granularities. The results obtained from reconstruction at different scales are weighted and summed to achieve comprehensive perception of multi-scale texture information.

[0088] In some embodiments, a multi-layered cascaded decoder decodes the enhanced features corresponding to each layer layer by layer to obtain the image processing result, including:

[0089] The multi-layered decoder decodes the enhanced features corresponding to each layer layer by layer to obtain multi-layered decoded features.

[0090] Among them, the decoded features are the features obtained through decoding operations (such as feature concatenation, upsampling, etc.).

[0091] Specifically, the input to each layer's decoder is the output of the previous layer's decoder, the corresponding enhancement features for that layer, and the image features for that layer obtained through skip connections. Of course, for the first layer's decoder, there is no output from the previous layer's decoder; therefore, the input is the corresponding layer's enhancement features and the corresponding layer's image features obtained through skip connections.

[0092] For the decoding features of each level, the offset is determined based on the decoding features, and offset convolution is performed based on the offset to obtain geometrically perceptual local features;

[0093] In some embodiments, the decoder includes independent convolutional layers that calculate the offset of each sampling point based on the decoding features. Then, the calculated offsets are used to adjust the position of the sampling points in the convolutional kernel, so that the convolutional kernel can adapt to the geometric deformation of different targets and obtain more accurate geometrically perceptual local features.

[0094] Shift window attention processing is performed based on the decoded features to obtain global attention features; the shift window attention processing includes window segmentation, cyclic shifting, self-attention calculation, and reverse shifting;

[0095] The shift window attention processing can be implemented by the shift window attention branch in the decoder. For example, the shift window attention branch first divides and rearranges the feature map of the decoded features according to the set window size to form an independent window sequence. Then, it performs a cyclic shift operation to move the window features cyclically in the specified direction so that the information of adjacent windows can interact. It calculates the attention weights of the query, key, and value vectors within the window and introduces relative position encoding to enhance position awareness. Finally, it restores the features to their original arrangement order through an inverse shift operation and rearranges them back to the feature map format to obtain the global attention features.

[0096] Based on geometric perception local features and global attention features, hierarchical fusion features are obtained.

[0097] For example, geometric perception local features and global attention features can be added together and fused to obtain hierarchical fused features.

[0098] Based on the fusion features at each level, the image processing results are obtained.

[0099] For example, the fusion features of each level are used as the output of the decoder of that level and as one of the inputs of the decoder of the next level. Finally, the images are decoded layer by layer through the decoders of multiple levels, and the image processing result is obtained based on the output of the decoder of the last level, that is, the fusion features of the last level.

[0100] In the above embodiments, after decoding the features, the decoder at each level extracts geometrically perceptual local features through offset convolution, which can improve the ability to model local structures and edge details. By obtaining global attention features through shift window attention, it is possible to model long-distance dependencies, capture global context information, and fuse geometrically perceptual local features with global attention features, taking into account both local accuracy and global semantic consistency, effectively improving feature expression ability and robustness, and ultimately making the image processing results more accurate and complete.

[0101] In some embodiments, image processing results are obtained based on the fusion features of each level, including:

[0102] Based on the fusion features of each level, intermediate features are determined.

[0103] In some embodiments, the fusion feature of the last level in each level can be determined as an intermediate feature.

[0104] In other embodiments, the fused features of each level may be upsampled to the same scale in sequence and spliced ​​in the channel dimension, and the spliced ​​features may be determined as intermediate features.

[0105] Spatial attention map is obtained by performing spatial attention weighting based on intermediate features.

[0106] Spatial attention weighting refers to capturing spatial dependencies in feature maps through depthwise separable convolution operations, generating a spatial attention map, which is used to weight pixel-level responses in the feature maps, highlighting foreground regions and suppressing background noise.

[0107] Channel attention maps are obtained by weighting channel attention based on intermediate features.

[0108] Channel attention weighting refers to aggregating global contextual information using global average pooling, then learning the importance of different channels through a multilayer perceptron structure to generate a channel attention map, which is used to filter feature channels that are more useful for the current task.

[0109] Image processing results are obtained based on spatial attention maps and channel attention maps.

[0110] In some embodiments, the spatial attention map and the channel attention map can be added together, fused with the original image features through residual connections, and then processed by an activation function to output the image processing result.

[0111] In the above embodiments, spatial attention weighting and channel attention weighting are applied to the intermediate features respectively, which can improve the feature expression ability from both spatial and channel dimensions, so that the final image processing result has higher accuracy, completeness and robustness.

[0112] In some embodiments, image processing results are obtained based on spatial attention maps and channel attention maps, including:

[0113] The spatial attention map and the channel attention map are added together to obtain the added features.

[0114] For example, the spatial attention map and the channel attention map can be added element by element to obtain the additive features.

[0115] The image processing results are obtained by fusing features based on additive features and image features of the initial synthetic aperture radar image.

[0116] For example, the additive features can be stitched or weighted and fused with the image features of the initial synthetic aperture radar image in the channel dimension. Then, the image processing results can be output by mapping and classification through operations such as activation and full connection.

[0117] In the above embodiments, by using the image features of spatial attention maps, channel attention maps, and initial synthetic aperture radar images, image details and spatial structure information can be restored while preserving global semantic information and key region information, thereby improving the accuracy and completeness of image processing results.

[0118] To facilitate understanding, a specific example application will be used below for illustration. Figure 5 As shown, Figure 5 A schematic diagram of another image processing system according to an embodiment of this application is shown; the image processing system includes a 4-layer cascaded encoder, a 4-layer cascaded decoder, and multiple wavelet decomposition modules.

[0119] The input synthetic aperture image is encoded by the first layer encoder. The output image features are used as input to the second layer encoder and as input to the corresponding wavelet decomposition module to achieve frequency domain enhancement. The output of the wavelet decomposition module (i.e. the enhanced features of the first layer) is used as input to the fourth layer decoder and also participates in the second-level processing through downsampling.

[0120] The second-layer encoder encodes the image features output by the first-layer encoder. These output features serve as input to the third-layer encoder and as input to the corresponding wavelet decomposition module to achieve frequency domain enhancement, resulting in the output of the wavelet decomposition module (i.e., the enhanced features of the second layer). Subsequently, the output of the wavelet decomposition module is concatenated with the downsampled features of the enhanced features of the first layer. The concatenated features serve as input to the third-layer decoder and, through downsampling, participate in the third-level processing.

[0121] The processing procedure for the third-layer encoder is similar to that for the second-layer encoder, and will not be described in detail here.

[0122] The fourth layer encoder is not connected to a wavelet decomposition module. The output of the fourth layer encoder does not need to be enhanced in the frequency domain by the corresponding wavelet decomposition module. Instead, the output of the fourth layer encoder is directly concatenated with the downsampled features of the enhanced features of the third layer, and the concatenated features are used as the input of the first layer decoder.

[0123] In some embodiments, if the image features extracted by the fourth layer encoder still contain a lot of detailed texture features, a corresponding wavelet decomposition module can be added between the fourth layer encoder and the first layer decoder.

[0124] Each wavelet decomposition module in each layer is as follows: Figure 6 As shown, Figure 6 The data processing flowchart of the wavelet decomposition module is shown.

[0125] The wavelet decomposition module includes wavelet bases at multiple scales, corresponding to convolution kernels at different scales. Figure 6The diagram shows different convolutional kernel sizes, including 2×2, 4×4, and 8×8. For each scale of the convolutional kernel, there are corresponding learnable parameters, which are the correction parameters of this application. The learnable parameters include position perturbation parameters and intensity scaling parameters. The wavelet base is adaptively adjusted through the learnable parameters. Its calculation logic is based on a linear combination of the basic wavelet base, position perturbation factor, and intensity scaling factor, so that the wavelet kernel can dynamically adjust its response characteristics according to the input features.

[0126] Among them, wavelet bases of multiple scales can be generated by bilinear interpolation based on an initial wavelet base. For example, interpolation is performed on an initial wavelet base of 2×2 to obtain wavelet bases of multiple scales, including 2×2, 4×4, and 8×8.

[0127] Each wavelet basis contains four frequency band components: LL (low-frequency approximation component), LH (horizontal edge component), HL (vertical edge component), and HH (diagonal edge component).

[0128] During wavelet decomposition, the module uses wavelet bases at each scale to perform discrete wavelet transform on the input features, obtaining low-frequency approximate components and three high-frequency detail components (horizontal, vertical, and diagonal), such as... Figure 7 As shown, Figure 7 The diagram shows the visualization of four frequency bands after discrete wavelet transform of the wavelet basis. It can be seen that the LL subband retains the main structure and background information, while the LH, HL, and HH subbands highlight the edge and texture details in different directions, which intuitively demonstrates the ability of frequency domain enhancement to capture edge information.

[0129] Furthermore, the processing results of the frequency bands can be compressed through channel attention. Then, since high-frequency components are crucial to the edges and textures in the synthetic aperture image, learnable weight parameters can be used to weight the processing results of different frequency bands, thereby filtering out the most discriminative feature information. At the same time, the wavelet decomposition module processes the wavelet basis transformation results at multiple scales. Since different scales correspond to different receptive fields and texture granularities, the wavelet decomposition module can use learnable scale weights to perform weighted summation of the results reconstructed at different scales, thereby achieving comprehensive perception of multi-scale texture information.

[0130] Please see Figure 8 , Figure 8 A schematic diagram illustrating the effect of frequency domain enhancement processing using the wavelet decomposition module provided in the embodiments of this application is shown.

[0131] Figure 8Four sets of effect illustrations are provided. For each set of effect illustrations, the top left corner of each image is the heatmap of the original image features before processing, the top right corner is the heatmap after frequency domain enhancement processing by the wavelet decomposition module, the bottom left corner is the original synthetic aperture image, and the bottom right corner is the effect of superimposing the original synthetic aperture image and the MS-Haar processed heatmap.

[0132] Depend on Figure 8 As can be seen, the heatmap after frequency domain enhancement processing by the wavelet decomposition module is shown. The background noise is significantly reduced, the small target points are clearer and brighter, and the texture and boundaries of different targets in the original synthetic aperture image can be fully reflected.

[0133] Each layer of the decoder contains a geometry-aware local-global attention module for offset convolution and shift window attention processing, such as... Figure 9 As shown, Figure 9 It shows Figure 5 A schematic diagram of the geometry-aware local-global attention module, including the geometry-aware local branch and the shifting window attention branch, wherein:

[0134] The geometry-aware local branch employs deformable convolution operations, generating offsets through independent convolutional layers and using these offsets to adjust the sampling point positions of the convolutional kernel, thereby adapting to the geometric deformation of the target. Due to the side-view imaging principle, synthetic aperture images often exhibit severe geometric distortions such as inverted tops and bottoms and overlapping shadows in man-made targets like buildings. Deformable convolution predicts the offset of each sampling point through additional convolutional layers. This allows the convolutional kernel to adaptively adjust the sampling position according to the actual shape of the target, such as sampling along sloping roof edges or deformed boundaries, thus capturing geometric structural features more accurately.

[0135] The shifted window attention branch divides the feature map into non-overlapping windows, calculates self-attention within each window, and enables information interaction between windows through a cyclic shift mechanism. The calculation process includes four steps: window segmentation, shifting, self-attention calculation, and reverse shifting. Specifically, the feature map is segmented and rearranged dimensionally according to a set window size to form an independent window sequence. Next, a cyclic shift operation is performed to move the window features cyclically in a specified direction, enabling information interaction between adjacent windows. Then, attention weights for query, key, and value vectors are calculated within the window, and relative position encoding is introduced to enhance position awareness. Finally, the reverse shift operation restores the features to their original arrangement order and rearranges them back into the feature map format.

[0136] Finally, the two feature paths are added and fused to obtain the output of the wavelet decomposition module. For example, the geometrically aware local features and the global attention features are aligned in the spatial dimension and then added and fused to combine the advantages of both. Subsequently, the fused features are projected and further processed through lightweight operations such as depthwise separable convolution to output an enhanced feature map.

[0137] In some embodiments, the decoder of the last layer further includes a feature refinement module, that is, the fourth layer decoder also includes a feature refinement module. The feature refinement module is used to perform a dual-path mechanism of spatial attention and channel attention in parallel on the fused features output by the decoder of the last layer.

[0138] The spatial attention path captures spatial dependencies in the feature map through depthwise separable convolution operations, generating a spatial attention map that is used to weight pixel-level responses in the feature map, highlighting the foreground region and suppressing background noise.

[0139] The channel attention path is used to aggregate global contextual information using global average pooling, and then learn the importance of different channels through a multilayer perceptron structure to generate a channel attention map, which is used to filter feature channels that are more useful for the current task.

[0140] Finally, the features weighted by the two attention streams are added together and fused with the original features through residual connections. After processing by an activation function, the refined features are output for the final classification prediction, thus obtaining the image processing result.

[0141] Please refer to Table 1 below, which shows a comparison of image segmentation performance data for different image processing methods.

[0142] Table 1

[0143]

[0144] Among them, UNet (U-shaped network), SegNet (segmentation network), ABCNet (adaptive boundary constraint network), SFA-Net (Spatial-Frequency Attention Network), UNetFormer (U-shaped Transformer network), and FSA-SARNet (frequency-guided spatial attention-network) are all image processing models in related technologies.

[0145] Per-class Intersection over Union (Per-class IoU) is used to reflect the segmentation accuracy for each class; a higher value indicates that the segmentation result for that class is more accurate.

[0146] Overall Accuracy (OA) reflects the proportion of all pixels that are correctly classified, and measures the overall classification accuracy.

[0147] The mean F1 score (mF1) is calculated by first calculating the F1 score (harmonic average of precision and recall) for each category, and then averaging it across all categories to reflect the balanced performance across all categories.

[0148] The mean Intersection over Union (mIoU) ratio is used to reflect the average segmentation accuracy of the model across all classes.

[0149] The number of parameters refers to the total number of trainable parameters in a model, reflecting the model's size and memory usage. Fewer parameters mean a more lightweight model.

[0150] Floating-point operations (FLOPs) refer to the number of floating-point operations required for a model to complete one forward inference operation, reflecting the model's computational complexity and inference speed. The smaller the FLOPs, the faster and more efficient the model's inference.

[0151] As can be seen from Table 1, the image processing system provided in this application embodiment has good recognition effect on various types of targets, and with fewer parameters and less computation, its various evaluation indicators are better than other related processing methods.

[0152] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0153] Based on the same inventive concept, this application also provides an image processing apparatus for implementing the image processing method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more image processing apparatus embodiments provided below can be found in the limitations of the image processing method described above, and will not be repeated here.

[0154] In one exemplary embodiment, such as Figure 10 As shown, a schematic diagram of an image processing apparatus is provided. This image processing apparatus is applied to an image processing system, which includes a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder. The multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection. The image processing apparatus 500 includes:

[0155] The first processing module 501 is used to acquire an initial synthetic aperture radar image, and to extract features from the initial synthetic aperture radar image layer by layer by a multi-layer cascaded encoder to obtain multi-level image features.

[0156] The second processing module 502 is used to perform multi-scale frequency domain enhancement processing and feature fusion based on the image features of each level by the wavelet decomposition module corresponding to that level, so as to obtain the fused features corresponding to that level; wherein the wavelet basis included in the wavelet decomposition module is determined based on the image features.

[0157] The third processing module 503 is used to perform layer-by-layer decoding and feature enhancement on the fusion features corresponding to each layer by the multi-layer cascaded decoder to obtain the image processing result.

[0158] In some embodiments, the image processing apparatus 500 further includes a construction module for constructing an initial wavelet basis; wherein the initial wavelet basis includes multiple different frequency band components; interpolation processing is performed based on the initial wavelet basis to obtain multiple wavelet bases at different scales; correction parameters are added to each wavelet basis to obtain a wavelet decomposition module; wherein the correction parameters include at least one of a position perturbation parameter and an intensity scaling parameter, and the correction parameters are used to adaptively adjust the wavelet basis according to image features.

[0159] In some embodiments, the second processing module 502 is configured to: perform parameter inference based on the image features of each level to obtain the target correction parameters corresponding to that level; deform the wavelet basis at multiple scales included in the wavelet decomposition module based on the target correction parameters to obtain the target wavelet basis at multiple scales; perform frequency domain decomposition at different scales on the image features of that level using the target wavelet basis at multiple scales to obtain the frequency domain features at different scales; and weight the frequency domain features at different scales to obtain the enhanced features corresponding to the level.

[0160] In some embodiments, the third processing module 503 is used to perform layer-by-layer decoding of the enhanced features corresponding to each layer by a multi-layer cascaded decoder to obtain multi-layer decoded features; for each layer's decoded features, determine the offset based on the decoded features, and perform offset convolution based on the offset to obtain geometrically perceptual local features; perform shift window attention processing based on the decoded features to obtain global attention features; wherein, the shift window attention processing includes window segmentation, cyclic shifting, self-attention calculation, and reverse shifting; obtain layer-by-layer fusion features based on the geometrically perceptual local features and global attention features; and obtain the image processing result based on the fusion features of each layer.

[0161] In some embodiments, the third processing module 503 is used to determine intermediate features based on the fusion features of each level; perform spatial attention weighting based on the intermediate features to obtain a spatial attention map; perform channel attention weighting based on the intermediate features to obtain a channel attention map; and obtain an image processing result based on the spatial attention map and the channel attention map.

[0162] In some embodiments, the third processing module 503 is used to add the spatial attention map and the channel attention map to obtain the added features; and to perform feature fusion based on the added features and the image features of the initial synthetic aperture radar image to obtain the image processing result.

[0163] Each module in the aforementioned image processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0164] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 11As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data related to image processing. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an image processing method.

[0165] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0166] In an exemplary embodiment, an image processing system is provided, comprising a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder; the multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection; wherein, the multi-layer cascaded encoder is used to extract features layer by layer from an initial synthetic aperture radar image to obtain multi-level image features; the wavelet decomposition module is used to perform multi-scale frequency domain enhancement processing and feature fusion based on the image features of each level, to obtain the fused features corresponding to that level; wherein the wavelet basis included in the wavelet decomposition module is determined based on the image features; the multi-layer cascaded decoder is used to perform layer-by-layer decoding and feature enhancement on the fused features corresponding to each level to obtain the image processing result.

[0167] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method provided in the above embodiments.

[0168] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the above embodiments.

[0169] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method provided in the above embodiments.

[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0171] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0172] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0173] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An image processing method, characterized in that, An image processing system, comprising a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder, wherein the multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection, the method comprising: An initial synthetic aperture radar image is acquired, and the multi-layer cascaded encoder performs layer-by-layer feature extraction on the initial synthetic aperture radar image to obtain multi-level image features; For each level of image features, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level; wherein, the wavelet basis included in the wavelet decomposition module is adapted to the image features; The multi-layered decoder decodes the enhanced features corresponding to each layer layer by layer to obtain the image processing result.

2. The method according to claim 1, characterized in that, The construction process of the wavelet decomposition module includes: Construct an initial wavelet basis; wherein the initial wavelet basis includes multiple different frequency band components; Interpolation is performed based on the initial wavelet basis to obtain multiple wavelet bases at different scales; The wavelet decomposition module is obtained by adding correction parameters to each wavelet basis; wherein the correction parameters include at least one of position perturbation parameters and intensity scaling parameters, and the correction parameters are used to adaptively adjust the wavelet basis according to the image features.

3. The method according to claim 2, characterized in that, For each level of image features, the wavelet decomposition module corresponding to that level performs multi-scale frequency domain enhancement processing based on the image features of that level to obtain the enhanced features corresponding to that level, including: For each level of image features, parameter inference is performed based on the image features of that level to obtain the target correction parameters corresponding to that level. Based on the target correction parameters, the wavelet bases at multiple scales included in the wavelet decomposition module corresponding to the level are deformed to obtain the target wavelet bases at multiple scales. Based on the target wavelet basis at multiple scales, frequency domain decomposition is performed at different scales for the image features at this level to obtain frequency domain features at different scales. The frequency domain features at different scales are weighted to obtain the enhanced features corresponding to the level.

4. The method according to claim 1, characterized in that, The image processing result is obtained by the multi-layer cascaded decoder performing layer-by-layer decoding on the enhanced features corresponding to each layer, including: The multi-layer cascaded decoder decodes the enhanced features corresponding to each layer layer by layer to obtain multi-layer decoded features; For each level of decoding features, an offset is determined based on the decoding features, and offset convolution is performed based on the offset to obtain geometrically aware local features; Based on the decoding features, shift window attention processing is performed to obtain global attention features; wherein, the shift window attention processing includes window segmentation, cyclic shifting, self-attention calculation, and reverse shifting; Based on the geometric perception local features and the global attention features, the fusion features of the level are obtained; Based on the fusion features of each of the aforementioned layers, the image processing results are obtained.

5. The method according to claim 1, characterized in that, The image processing result obtained based on the fusion features of each of the aforementioned levels includes: Based on the fusion features of each of the aforementioned levels, intermediate features are determined; Spatial attention weighting is performed based on the intermediate features to obtain a spatial attention map; Channel attention map is obtained by performing channel attention weighting based on the intermediate features. Based on the spatial attention map and the channel attention map, the image processing result is obtained.

6. The method according to claim 5, characterized in that, The image processing result obtained based on the spatial attention map and the channel attention map includes: The spatial attention map and the channel attention map are added together to obtain the additive features; Based on the additive features and the image features of the initial synthetic aperture radar image, feature fusion is performed to obtain the image processing result.

7. An image processing apparatus, characterized in that, An image processing system, comprising a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder, wherein the multi-layer cascaded encoder and the multi-layer cascaded decoder are interconnected, the device comprising: The first processing module is used to acquire an initial synthetic aperture radar image, and the multi-layer cascaded encoder performs layer-by-layer feature extraction on the initial synthetic aperture radar image to obtain multi-level image features. The second processing module is used to perform multi-scale frequency domain enhancement processing and feature fusion based on the image features of each level by the wavelet decomposition module corresponding to that level, to obtain the fused features corresponding to that level; wherein, the wavelet basis included in the wavelet decomposition module is determined based on the image features; The third processing module is used to perform layer-by-layer decoding and feature enhancement on the fusion features corresponding to each of the multi-layer cascaded decoders to obtain the image processing result.

8. An image processing system, characterized in that, The image processing system includes a multi-layer cascaded encoder, a wavelet decomposition module, and a multi-layer cascaded decoder; the multi-layer cascaded encoder and the multi-layer cascaded decoder are connected in a skip connection, wherein: The multi-layer cascaded encoder is used to extract features layer by layer from the initial synthetic aperture radar image to obtain multi-level image features; The wavelet decomposition module is used to perform multi-scale frequency domain enhancement processing and feature fusion based on the image features of each level, thereby obtaining the fused features corresponding to that level; wherein, the wavelet basis included in the wavelet decomposition module is determined based on the image features; The multi-layer cascaded decoder is used to perform layer-by-layer decoding and feature enhancement on the fusion features corresponding to each layer to obtain the image processing result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images