Multi-kernel SAR target detection method based on adaptive convolution

Abstract

The invention provides a multi-kernel SAR target detection method based on adaptive convolution, which comprises the following steps: firstly, receiving an input SAR image, sending the SAR image into a feature extraction network, after each convolution layer, using a SiLU activation function to introduce nonlinear transformation and enhance the expression ability of a model, and finally outputting an accurate target detection result. Comprising detailed information such as position, category and direction of the target. According to the method, the limitation of a traditional SAR target detection method in a complex environment is effectively solved, and the target detection performance is remarkably improved. According to the method, the challenge of multi-scale target detection is effectively dealt with, the accuracy of target recognition under a complex background is improved, the robustness of the model is remarkably enhanced, the performance of SAR image target detection is greatly improved, and the method has important value and wide application prospects in practical applications such as military reconnaissance, disaster monitoring and earth observation.

Description

Technical Field

[0001] The invention relates to the technical field of synthetic aperture radar (SAR) target detection, and in particular to a SAR image ship target detection method, which can be used for ship target detection tasks in complex environments. Background Art

[0002] Synthetic aperture radar (SAR) technology, with its excellent high-resolution imaging characteristics, occupies an important position in the contemporary military and civilian fields. This technology captures the image information of the target by emitting and receiving the reflected electromagnetic waves. Its application scope covers many key fields such as earth observation, military reconnaissance, and disaster monitoring. However, the acquisition process of SAR images is not without challenges. They are often interfered by complex environmental factors such as atmospheric disturbances, multipath effects, and surface clutter. These interference factors may mask the target signal or make it blurred, thereby increasing the difficulty of target detection. In addition, the accurate identification of the target depends not only on its appearance characteristics, but also on its contextual environment. For example, a ship may be located at sea or close to the shore. The multi-scale characteristics of the target in the SAR image, that is, the target appears at different sizes, adds additional complexity to the design of the detection algorithm.

[0003] In order to cope with the challenges brought by the change of target scale in SAR images, researchers have adopted a series of strategies to improve the robustness of detection algorithms to scale changes. One of them is the explicit data augmentation technique, which simulates targets of different scales in the training set to enhance the model's adaptability to scale changes. At the same time, the multi-scale feature integration method aims to provide more comprehensive scale information by integrating features extracted at different resolutions. The pyramid feature hierarchy captures target information of different scales by constructing a multi-level feature representation.

[0004] However, existing methods still have limitations in processing SAR images. A core problem is that objects of different scales may require receptive fields of different sizes to effectively capture their contextual information. Currently, many methods use fixed-size receptive fields to process objects of different scales, which limits the network's ability to capture contextual information of large objects and affects the overall performance of SAR image object detection.

[0005] Existing SAR image target detection methods attempt to address the challenge by expanding the spatial receptive field of the background network, but this approach has some limitations. For example, the use of large convolution kernels may introduce additional background noise, which may have an adverse effect on the accurate detection of small targets. This is because large convolution kernels may capture noise signals that are irrelevant to the target when extracting background features, thereby interfering with the feature representation of small targets. Therefore, how to reduce the potential loss of small target detection accuracy while increasing context information has become an unsolved problem. In addition, although dilated convolution can effectively expand the receptive field, if a larger dilation rate is used, it may reduce the response to high-frequency signals, resulting in the failure to capture high-frequency detail information in the image. This situation is particularly prominent when processing images with rich details, which may lead to a decrease in detection performance. Traditional dilated convolution methods use a globally fixed dilation rate, which does not fully consider the local frequency characteristics of different regions in the image and may not be able to optimally adapt to all parts of the image. Summary of the invention

[0006] In order to overcome the shortcomings of the prior art, the present invention provides a multi-core SAR target detection method based on adaptive convolution. Multi-scale convolution kernels and adaptive dilated convolutions are used to extract target features of different scales and capture local context. The present invention designs a convolutional neural network structure with the ability to dynamically adjust the receptive field so as to better capture and understand the contextual information of targets of different scales. By combining these two structures, the accuracy and robustness of target detection in SAR images can be significantly improved. The technical solution of the present invention mainly includes two parts: a multi-core context-aware network and a spectrum adaptive dilation network. In the proposed multi-core synthetic aperture radar (SAR) target detection method based on adaptive convolution, an input SAR image is first received and sent to a feature extraction network. This feature extraction network includes multiple convolutional layers and batch normalization layers for extracting features from the input image. After each convolutional layer, a SiLU activation function is used to introduce nonlinear transformations and enhance the expressive power of the model.

[0007] The multi-core context-aware network adopts a parallel processing method and uses convolution kernels of different sizes to capture multi-scale target features. In addition, through the context center anchor attention module, the network can effectively capture long-distance context information, thereby improving the accuracy of detection. The present invention introduces a spectrum adaptive expansion structure, which dynamically adjusts the expansion rate of the convolution through spectrum analysis to optimize the balance between the receptive field and the effective bandwidth. The introduction of this structure enables the model to better adapt to targets of different scales. The spatial pyramid fast pooling (Spatial PyramidPooling, SPP) technology is used to further enhance the model's perception of spatial information. By performing pooling operations of different scales on the input feature map, the SPP technology enables the model to effectively capture information of different scales in the image. In addition, the progressive asymmetric feature pyramid network is used to construct a feature pyramid, and the detection performance of targets of different sizes is improved through multi-scale fusion. The design of this network structure enables the model to more comprehensively consider the characteristics of the target at different scales. The detection head is designed to be decoupled, which means that it can be optimized independently of the feature extraction network. This design makes the training process of the model more flexible and can be customized and optimized for specific tasks.

[0008] Finally, the framework can output accurate target detection results, including detailed information such as the location, category, and direction of the target. Through this comprehensive design, the method can effectively handle multi-scale targets in the SAR image target detection task, while fully considering the shape and context information of the target, achieving high-performance detection results.

[0009] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0010] Step 1: Dataset preparation;

[0011] The network structure is trained and tested using the HRSID public dataset, which is designed to handle ship detection, semantic segmentation, and instance segmentation tasks in high-resolution synthetic aperture radar (SAR) images. The HRSID public dataset has been widely used in the development and benchmarking of SAR image analysis algorithms, especially in ship detection. It contains 5,604 high-resolution SAR images and annotations of 16,951 ship instances. The SAR images are from Sentinel-1B, TerraSAR-X, and Tandemi-X sensors, ensuring coverage of various imaging conditions and resolutions. The HRSID dataset includes high-resolution images and captured diverse scenes (including different sea conditions, sea areas, and coastal ports), of which 3,642 are used for training and 1,962 are used for testing. The image size is 800×800.

[0012] Step 2: Construction of multi-core SAR target detection model based on adaptive convolution;

[0013] The target detection model includes convolution blocks, multi-core context-aware networks, spectrum adaptive expansion structures, progressive asymmetric feature pyramid networks, and decoupled detection heads. After the SAR image is input, it passes through the convolution blocks and then is input into the multi-core context-aware network. The output of the multi-core context-aware network is sequentially processed through the spectrum adaptive expansion structure and spatial pyramid fast pooling. The image that has passed through the multi-core context-aware network and spatial pyramid fast pooling is input into the progressive asymmetric feature pyramid network for processing to obtain a feature map, which is then input into the decoupled detection head for post-processing.

[0014] Step 3: Loss function;

[0015] Step 4: Use loss function The network model is trained using the training set and the validation set is used during the training process to observe the model performance and obtain the network model with the highest map50 and map50:95 indicators.

[0016] In step 2, the construction process of the target detection model is as follows:

[0017] Step 2-1: Construct a convolution block, and connect three convolution blocks in series;

[0018] The convolution block includes convolution layer, batch normalization and Silu activation function. The convolution layer window size is 3×3, the padding is 1, and the stride is 1.

[0019] Step 2-2: Build a multi-core context-aware network;

[0020] The multi-core context-aware network is composed of three multi-core context-aware structures connected in series;

[0021] Step 2-3: Construct a spectrum adaptive expansion structure;

[0022] The spectrum adaptive dilation structure is designed to process highly variable spatial frequency information and is divided into three functional blocks: adaptive dilation rate module, frequency selection module and adaptive kernel module; the frequency selection module analyzes and optimizes the performance of the input feature map at different frequencies, adjusts the contribution of different frequency bands by adjusting the weights of the frequency components, and outputs a weighted synthetic feature map; the adaptive kernel module adjusts the components of the convolution kernel to better capture high-frequency and low-frequency information, and outputs an adaptive convolution kernel; the feature map output by the frequency selection module and the adaptive convolution kernel output by the adaptive kernel module are the inputs of the adaptive dilation rate module. First, a dilation map is generated according to the local characteristics of the input feature map. Each value in the dilation map defines the dilation rate that should be used for the dilated convolution at the corresponding position. According to the dilation rate indicated by the dilation map, the dilated convolution is applied to the input feature map. The dilated convolution increases its receptive field by inserting spaces (determined by the dilation rate) between the elements of the convolution kernel without increasing the number of parameters or computational complexity. The final output of the adaptive dilation structure is a feature map that has undergone adaptive dilation convolution;

[0023] Step 2-4: Construct spatial pyramid fast pooling, progressive asymmetric feature pyramid network and decoupled detection head;

[0024] The input of the spatial pyramid fast pooling is the F4 output of the spectrum adaptive expansion structure with 256 channels, and the output channel number is 256. The input of the progressive asymmetric feature pyramid network comes from the feature map F2 output by the multi-core context-aware structure 2, the F3 output by the multi-core context-aware structure 2, and the output F5 of the spatial pyramid fast pooling, with the channel numbers of 64, 128, and 256 respectively. The three feature maps of sizes 20*20, 40*40, and 80*80 output by the progressive asymmetric feature pyramid network are used as the input of the three decoupled detection heads.

[0025] Step 2-5: Decoupling detection;

[0026] The detection head uses the three feature maps of 20*20, 40*40, and 80*80 sizes output by the progressive asymmetric feature pyramid network as the input of the three decoupled detection heads. Traditional detection heads often use the same set of features to simultaneously complete the classification and regression tasks of the target. However, due to the different requirements of the classification and regression tasks, this way of sharing features may cause conflicts between the two, thereby affecting the detection accuracy; the decoupled detection head separates the classification and regression tasks and uses different branches to handle these two tasks respectively, thereby improving the detection performance; the classification branch focuses on the classification task of the target, inputs the feature map output by the feature pyramid network, and extracts the feature map for The features of the classification are used to predict the target category score in each detection frame. The output shape of the classification branch is a four-dimensional vector, which includes the batch size, the number of anchors * the number of categories, the width of the feature map, and the height of the feature map. The regression branch focuses on the regression task of the bounding box, that is, predicting the location of the target. Similarly, the input is the feature map output by the progressive asymmetric feature pyramid network, and the features for regression are extracted through a series of different convolutional layers. The features output by the regression branch are used to predict the location information of each bounding box. The location information includes the coordinates of the upper left corner and the lower right corner of the bounding box, or the center point and the width and height. The output shape is the batch size, the number of anchors * 4, the width of the feature map, and the height of the feature map.

[0027] Step 2-6: Post-processing;

[0028] After the decoupled detection head outputs the detection result, it needs to go through post-processing steps to obtain the final target detection result. The post-processing steps include bounding box decoding, non-maximum suppression, threshold filtering and result output. The four steps are performed in series. Bounding box decoding decodes the position information of the detection box predicted by the regression branch into actual coordinates relative to the original image; for example, the regression head outputs the center point offset and aspect ratio, which are converted into the actual upper left corner and lower right corner coordinates; then the non-maximum suppression operation is performed. First, the bounding box with the highest score is selected as the current box, and then other boxes with an overlap (IoU) exceeding 0.5 are removed, and the non-maximum suppression process is repeated in sequence until there are no candidate boxes that meet the conditions; finally, the bounding boxes with a confidence threshold lower than 0.001 are removed, and the retained boxes are the detection results;

[0029] The loss function described in step 3 includes category classification loss and bounding box regression loss:

[0030] The classification loss uses the cross entropy loss Binary Cross-Entropy Loss to measure the model's prediction accuracy for the target category, as shown in formula (9):

[0031]

[0032] Where N is the number of bounding boxes, y i is the true label, 1 means the target exists, 0 means the target does not exist, p i is the confidence score of the model prediction;

[0033] The bounding box regression loss is in the form of Generalized Intersection over Union (CIOU Loss) and Distance-based Feature Learning Loss (DFL Loss);

[0034] DFL Loss is used to optimize the bounding box regression task. It does not directly predict the coordinates of the bounding box, but predicts the distribution of the coordinates (for example, through a set of basis functions), so as to more accurately estimate the position of the bounding box through the combination of basis functions. The FL formula is as shown in formula (10):

[0035] DFL(S i ,S i+1 )=-(y i+1 -y)log(S i )+(yy i )log(S i+1 ) (10)

[0036] Here, S i and S i+1 Represent the probability of the predicted distribution in the i-th interval and the i+1-th interval respectively, and y is the true label or value; the loss function improves the prediction accuracy of the model by optimizing the distance between the predicted distribution and the true value;

[0037] CIoU Loss is a loss function for bounding box regression that takes into account the overlapping area, center point distance, and aspect ratio between bounding boxes, thus providing a more comprehensive loss evaluation; as shown in formula (11):

[0038]

[0039] Among them, ρ(b,b gt ) represents the predicted box b gt The distance between the real box b, c is the diagonal length of the smallest closed box containing the two boxes, v and α are terms related to the aspect ratio, which are used to adjust the contribution of the aspect ratio contrast loss;

[0040] The overall loss function As shown in formula (12):

[0041]

[0042] In step 2-2, in the multi-core context-aware structure, the input is the feature map F from the previous stage l-1 , l represents the current layer l of the multi-core context-aware structure, l-1 represents the previous layer, the input feature map is first reduced in spatial size by the Haar wavelet downsampling module; the 3×3 convolution layer performs a convolution operation on the feature map after Haar wavelet downsampling to extract semantic features, where the 3×3 convolution layer changes the depth of the feature map, that is, the number of channels; the feature map X after the 3×3 convolution layer is l-1 Divide into two parts evenly according to the number of channels and The two parts are then assigned to two different processing paths; one path is a feed-forward network (FFN), which processes the input feature map through two convolutional layers and activation functions in series. Perform direct forward propagation to extract and transfer feature information, and the output is The other path consists of multiple multi-core context-aware layers connected in series, each of which is a basic building block in the network, and the output is

[0043] The output of the feed-forward network FFN Output of the multi-core context-aware layer Then, the convolutional layer is used to further integrate the features and output the final feature map F. l , the output F of the entire multi-core context-aware structure l The expression of is shown in formula (1):

[0044]

[0045] Wherein, Concat represents a concatenation operation, and Conv represents a convolution operation. The concatenation and convolution operations help to integrate the features of different processing paths and enhance the representation capability of the model of the present invention for input data.

[0046] Multi-core context-aware networks such as Figure 3 As shown, the multi-core context-aware structure is as follows Figure 4 As shown in Figure 2, the multi-core context perception layer integrates the multi-core context perception module and the context center anchor attention module, and the two modules are performed in parallel. Figure 5 As shown, the input of the multi-core context perception layer is the feature map passed in from the previous structure or the previous multi-core context perception layer The multi-core context perception module uses multi-scale convolution kernels to capture target features of different scales; the input of the context center anchor attention module is also the feature map passed in by the previous structure or the previous multi-core context perception layer. Focus on capturing the contextual relationship between distant pixels in the image; the context center anchor attention module first obtains global features through average pooling, and then extracts and refines the global features through a 1×1 convolution layer and a 3×3 depth-separable convolution, thereby enhancing the semantic information of the features; the feature maps output by the multi-core context perception module and the context center anchor attention module are multiplied, and the result of the multiplication is added to the feature map output by the multi-core context perception module. The fused feature map is processed through a 1x1 convolution layer to further integrate the features and reduce the feature dimension, and the output is The processing results of the two parallel paths are finally combined through channel concatenation and fusion operations to generate the output feature maps of each stage;

[0047] In the multi-core context-aware module, the input of the multi-core context-aware module is The feature map output after processing by the multi-core context-aware module is P l-1,n , used as input for subsequent layers or modules; the initial depth-separable convolution operation uses a 3x3 convolution kernel to align the input Feature extraction is performed, and then multi-scale depthwise separable convolutions are performed using 5x5, 7x7, 9x9, and 11x11 convolution kernels. These operations help the model capture spatial features of different scales. Identity mapping is used to directly pass the 3×3 depthwise separable convolution to the fusion stage and fuse with feature maps of other scales.

[0048] In the context center anchor attention module, the formula of the context center anchor attention module is expressed as formula (2):

[0049]

[0050] Average pooling operation (AvgPool) on the input feature map Downsampling of the spatial dimension is performed to reduce the size of the feature map to reduce the amount of subsequent calculations. Average pooling is achieved by calculating the average value of each small area in the input feature map. Then a 1x1 convolution layer is used to adjust the number of channels. This step will not change the spatial size of the feature map and is mainly used to linearly combine the features of different channels. The first step is a 1x(11+2xn) depthwise separable convolution to process the features in the width direction. This is followed by a (11+2xn)x1 depthwise separable convolution to process the features in the height direction. The depthwise separable convolution performs spatial convolution on each input channel separately, and then applies a 1x1 convolution to fuse the channel information. The features after the depthwise convolution are integrated again through a 1x1 convolution, which further improves the expressiveness of the features and the ability to integrate features of different sizes. Finally, the final convolution output is processed through the Sigmoid activation function to obtain the final attention weight map A. l-1,n, the Sigmoid function compresses the output value between 0 and 1, indicating the importance of each position.

[0051] In step 2.3, the steps of constructing the spectrum adaptive expansion structure are:

[0052] Step 2-3-1: Build a frequency selection module;

[0053] In the frequency selection module, the input spatial domain feature map is the output F3 of the multi-core context-aware network, which is first converted to the frequency domain through Fourier transform to form X F , the transformation reveals the different frequency components in the image; then, X F is decomposed into a number of predefined frequency bands, each of which is processed by applying a corresponding frequency domain mask M b Extract the features of the corresponding frequency range Each frequency band feature is independently passed through the convolution kernel M designed for it b Processing is performed to generate the convolution result Y b To determine the contribution of each frequency band to the final output, b Apply 1×1 convolution and Sigmoid activation function to generate selection map S b , select Figure S b Expresses the importance of each frequency band; ultimately, all weighted frequency band features Y b By comparing their respective selections to b Multiply and accumulate to synthesize the final output feature map Y, so that the module can dynamically adjust the focus on different frequency bands and optimize the overall feature extraction and information synthesis. The final output feature map is calculated by formula (3):

[0054] Y=∑ b S b ·Y b (3)

[0055] Among them, S b =Sigmoid(Conv(X b )) represents the selection weight obtained by applying 1x1 convolution and Sigmoid function to each frequency band feature. The re-weighted frequency band features are weighted summed to form the final output feature map, which has optimized frequency components and is more suitable for complex spatial processing;

[0056] Step 2-3-2: Build the adaptive selection kernel module;

[0057] The input of the adaptive selection kernel module is the output F3 of the multi-core context-aware network, Global Pool represents the global pooling operation, Conv represents the convolution operation, ReLU is the nonlinear activation function, and σ is the Sigmoid function, which is used to compress the output of the convolution operation to between 0 and 1; as shown in formula (4):

[0058] G=σ(Conv(ReLU(Conv(local Pool(F3))))) (4)

[0059] The adaptive selection kernel module uses global pooling to obtain the global context information of the input feature map, which is further processed through a series of convolution and ReLU layers, and then a regulation factor G is generated through the Sigmoid activation function;

[0060] The initial convolution kernel K is decomposed into low-frequency K l and high frequency K h The two parts dynamically adjust the combination of the two kernels according to the global feature map adjustment factor G, as shown in formula (5):

[0061] K′=G·K l +·K(1-G)·K h (5)

[0062] K l =Average(K) (6)

[0063] K h =KK l (7)

[0064] Among them, K′ is the final adaptive convolution kernel, which is used to process the input features, G is the adjustment factor extracted from the global features, which controls the influence of the low-frequency kernel, and 1-G naturally controls the influence of the high-frequency kernel to ensure the balance of low-frequency and high-frequency components; Average is the mean value;

[0065] The core of the adaptive kernel selection module is how to use the global feature map to balance the low-frequency K l and high frequency K h The kernel contribution,G,value determines the relative importance of low-frequency information and high-frequency information,in a specific image region;

[0066] Step 2-3-3: Construct an adaptive expansion rate selection module;

[0067] The whole process of adaptively selecting the expansion rate module is shown in formula (8):

[0068] Y(p)=∑ k∈K X(p+r(p)·K′)·W(K′) (8)

[0069] Among them, K′ is the adaptive convolution kernel output by the adaptive selection kernel module, X(p) is the value of the input feature map at position p, W(K′) is the weight of the convolution kernel at position k, r(p) is the dilation rate of the dilation map at position p, which is dynamically generated by the previous convolution + ReLU layer, and Y(p) is the feature map finally output by the spectrum adaptive dilation structure.

[0070] An electronic device comprises one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to execute the method described above.

[0071] A computer-readable storage medium stores program codes, and the program codes can be called by a processor to execute the method as described above.

[0072] The beneficial effect of the present invention is that the proposed multi-core SAR target detection method based on adaptive dilated convolution adopts technical means such as multi-core multi-scale and frequency-domain-based adaptive dilated convolution, which effectively solves the limitations of traditional SAR target detection methods in complex environments and significantly improves the performance of target detection. This method not only effectively responds to the challenges of multi-scale target detection, but also improves the accuracy of target recognition under complex backgrounds and significantly enhances the robustness of the model. Ultimately, the solution provided by the present invention greatly improves the performance of SAR image target detection, and has important value and broad application prospects in practical applications such as military reconnaissance, disaster monitoring, and earth observation. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of the method of the present invention.

[0074] Figure 2 It is a structural diagram of the network model constructed by the present invention.

[0075] Figure 3 It is a schematic diagram of a multi-core context-aware network of the present invention.

[0076] Figure 4 It is a schematic diagram of the multi-core context-aware structure of the present invention.

[0077] Figure 5 It is a schematic diagram of the multi-core context perception layer of the present invention.

[0078] Figure 6 It is a schematic diagram of the multi-core context perception module of the present invention.

[0079] Figure 7 It is a schematic diagram of the spectrum adaptive expansion structure of the present invention.

[0080] Figure 8It is a comparison chart of the detection results of the existing method and the method of the present invention in the embodiment of the present invention. DETAILED DESCRIPTION

[0081] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0082] The present invention designs a convolutional neural network structure with the ability to dynamically adjust the receptive field in order to better capture and understand the contextual information of targets of different scales. By combining these two structures, the accuracy and robustness of target detection in SAR images can be significantly improved. The technical solution of the present invention mainly consists of two parts: a multi-core context-aware network and a spectrum adaptive expansion network. Figure 1 As shown in the figure, in the proposed multi-core SAR target detection method based on adaptive convolution, the input SAR image is first received and fed into a feature extraction network, which contains multiple convolutional layers and batch normalization layers. These layers are followed by SiLU activation functions to introduce nonlinearity and enhance the expressiveness of the model. The core network, i.e., the multi-core context-aware network, uses convolution kernels of different sizes for parallel processing to capture multi-scale target features, and captures long-range context information through the context center anchor attention module, thereby improving the accuracy of detection. In addition, the present invention adopts a spectrum adaptive expansion structure, which dynamically adjusts the expansion rate of the convolution through spectrum analysis to optimize the balance between the receptive field and the effective bandwidth. The spatial pyramid fast pooling (Spatial Pyramid Pooling, SPP) technology further enhances the model's perception of spatial information. The progressive asymmetric feature pyramid network is used to construct a feature pyramid, and the detection performance of targets of different sizes is improved through multi-scale fusion. The detection head is designed to be decoupled, allowing optimization independent of the feature extraction network.

[0083] Finally, the framework can output accurate target detection results, including detailed information such as the location, category, and direction of the target. Through this comprehensive design, the method can effectively handle multi-scale targets in the SAR image target detection task, while taking into account the shape and context information of the target, achieving high-performance detection results.

[0084] The specific steps include:

[0085] Step 1: Dataset preparation;

[0086] The public dataset HRSID is used for training and testing the network structure of the present invention. HRSID is designed for ship detection, semantic segmentation, and instance segmentation tasks using high-resolution synthetic aperture radar (SAR) images. This dataset is widely used in the development and benchmarking of SAR image analysis algorithms, especially for detecting ships. The dataset contains 5,604 high-resolution SAR images and annotations of 16,951 ship instances. These images were acquired from Sentinel-1B, TerraSAR-X, and Tandemi-X sensors, ensuring a variety of imaging conditions and resolutions. The HRSID dataset is known for its high-resolution images and the different scenes it captures (including different sea conditions, sea areas, and coastal ports), of which 3,642 are training sets and 1,962 are test sets.

[0087] Step 2: Construction of multi-core SAR target detection model based on adaptive convolution;

[0088] The network model is divided into: building convolution blocks, multi-core context-aware networks, spectrum adaptive expansion structures, progressive asymmetric feature pyramid networks, and decoupled detection heads.

[0089] The specific construction process of each part is as follows:

[0090] Step 2-1: Construct a convolution block;

[0091] The convolution block is mainly composed of convolution layer, batch normalization, and Silu activation function. The convolution layer window size is 3×3, the padding is 1, and the step size is 1;

[0092] Step 2-2: Build a multi-core context-aware network;

[0093] like Figure 3 As shown, the multi-core context-aware network is composed of three multi-core context-aware structures in sequence.

[0094] (1) Multi-core context-aware architecture

[0095] Each stage adopts a cross-stage partial (CSP) structure, which allows efficient feature processing and optimization of computational load. The multi-core context-aware architecture adopts this structure, such as Figure 4 In the multi-core context-aware structure shown in Figure 2, the input is the feature map F from the previous stage. l-1 The input feature map is firstly processed by the Haar wavelet downsampling module to reduce the spatial size, reduce the amount of calculation and retain the semantic information of the target as much as possible. The convolution module performs convolution operation on the downsampled feature map to extract the semantic features. This step will change the depth of the feature map, that is, the number of channels. l-1Strategically split into two parts and Then it is assigned to two different processing paths. One path is a feed-forward network (FFN), which directly forward propagates the input features through a series of convolutional layers and activation functions to extract and transmit feature information, and the output is The other path is more complex and consists of multiple multi-core context-aware layers. Each multi-core context-aware block is a basic building block in the network, responsible for further refining and enhancing the feature representation, and the output is

[0096] The output of FFN Output of the multi-core context-aware layer Then, the convolutional layer is used to further integrate the features and output the final feature map F. l The output F of the entire multi-core context-aware structure is l The expression of is shown in formula (1).

[0097]

[0098] Wherein, Concat represents a concatenation operation, and Conv represents a convolution operation. The concatenation and convolution operations help to integrate the features of different processing paths and enhance the representation capability of the model of the present invention for input data. l represents the current layer l of the multi-core context-aware structure;

[0099] (2) Multi-core Context Perception Layer

[0100] like Figure 5 As shown in Figure 1, each multi-core context perception layer integrates two key modules: the multi-core context perception module and the context center anchor attention module. The input of the multi-core context perception layer is the feature map passed in from the previous structure or the previous multi-core context perception layer. The multi-core context-aware module uses multi-scale convolution kernels to capture target features of different scales, while the context center anchor attention module focuses on capturing the contextual relationship between distant pixels in the image, thereby enhancing the semantic information of the features. The fused feature map is processed through a 1x1 convolution layer to further integrate the features and reduce the feature dimension, outputting

[0101] The processing results of these two parallel paths are finally combined through channel cascading and fusion operations to generate the output feature map of each stage. This design not only improves the network's ability to detect multi-scale targets, but also significantly enhances its adaptability to complex contextual environments. In this way, the network can effectively extract features with rich local and global contextual information, providing strong support for the accurate detection of targets in SAR images.

[0102] (3) Multi-core context-aware module

[0103] like Figure 6 As shown, the input of the multi-core context-aware module is The feature map output after processing by this module is P l-1,n , used as input for subsequent layers or modules. The initial depth-wise separable convolution operation uses a 3x3 convolution kernel to align the input Feature extraction is performed. Then, multi-scale deep separable convolution uses 5x5, 7x7, 9x9 and 11x11 convolution kernels to perform deep separable convolution operations. These operations help the model capture spatial features of different scales. Identity mapping is used to pass the 3×3 deep separable convolution directly to the fusion stage and fuse it with feature maps of other scales. 1x1 convolution is used for all feature maps, which helps to adjust the channel weights of different feature maps and enhance the expressiveness of the model.

[0104] (4) Contextual Center Anchor Attention Module

[0105] The main function of the contextual center anchor attention module is to strengthen the feature map through the contextual anchor attention mechanism, so that the model can focus on important information in the remote context. This is particularly important in SAR image processing because objects in the image may show correlations over a very wide area. In this way, the contextual center anchor attention module helps the present invention to more effectively process SAR images with complex backgrounds and different scale changes, improving the accuracy and robustness of the detection task.

[0106] The key formula of this module is expressed as formula (2):

[0107]

[0108] Average pooling operation (AvgPool) on the input feature map Downsample the spatial dimension to reduce the size of the feature map to reduce the amount of subsequent calculations. Average pooling is achieved by calculating the average value of each small area in the input feature map. Then use a 1x1 convolution layer to adjust the number of channels. This step will not change the spatial size of the feature map and is mainly used to linearly combine the features of different channels. The first is a 1x(11+2xn) depthwise separable convolution to process features in the width direction. This is followed by a (11+2xn)x1 depthwise separable convolution to process features in the height direction. The depthwise separable convolution performs spatial convolution on each input channel separately, and then applies a 1x1 convolution to fuse the channel information. This step again integrates the features after the depthwise convolution through a 1x1 convolution, further improving the expressiveness of the features and the ability to integrate features of different sizes. Finally, the final convolution output is processed by the Sigmoid activation function to obtain the final attention weight map A. l-1,n The Sigmoid function compresses the output value between 0 and 1, indicating the importance of each position.

[0109] Step 2-3 constructs a spectrum adaptive expansion structure;

[0110] The spectrum adaptive dilation structure is designed for processing highly variable spatial frequency information. It is mainly divided into three functional blocks: adaptive dilation rate module, frequency selection module, and adaptive kernel module. The frequency selection module analyzes and optimizes the performance of the input feature map at different frequencies, adjusts the contribution of different frequency bands by adjusting the weights of the frequency components, and outputs a weighted synthetic feature map; the adaptive kernel module adjusts the components of the convolution kernel to better capture high-frequency and low-frequency information, and outputs an adaptive convolution kernel; the feature map output by the frequency selection module and the adaptive convolution kernel output by the adaptive kernel module are the inputs of the adaptive dilation rate module. First, a dilation map is generated according to the local characteristics of the input feature map. Each value in the dilation map defines the dilation rate that should be used for the dilated convolution at the corresponding position. According to the dilation rate indicated by the dilation map, the dilated convolution is applied to the input feature map. The dilated convolution increases its receptive field by inserting spaces (determined by the dilation rate) between the elements of the convolution kernel without increasing the number of parameters or computational complexity. The final output of the adaptive dilation structure is a feature map after adaptive dilation convolution.

[0111] Step 2-3-1 build frequency selection module;

[0112] In the workflow of the frequency selection module, the input spatial domain feature map is the output F3 of the multi-core context-aware network, which is first converted to the frequency domain through Fourier transform to form X F This transformation reveals the different frequency components in the image. Next, X Fis decomposed into a number of predefined frequency bands, each of which is processed by applying a corresponding frequency domain mask M b To extract the features of the corresponding frequency range Each frequency band feature is then independently convolved through a specific convolution kernel W designed for it b Processing is performed to generate the convolution result Y b To determine the contribution of each frequency band to the final output, Y b Apply 1×1 convolution and Sigmoid activation function to generate selection map S b , this selection graph expresses the importance of each frequency band. Finally, all weighted frequency band features Y b By comparing their respective selections to b Multiply and accumulate to synthesize the final output feature map Y. This process enables the module to dynamically adjust its focus on different frequency bands and optimize the overall feature extraction and information synthesis. The final output feature map is calculated by formula (3):

[0113] Y=∑ b S b ·X b (3)

[0114] Among them, S b =Sigmoid(Conv(X b )), represents the selection weight obtained by applying 1x1 convolution and sigmoid function to each frequency band feature. The reweighted frequency band features are combined (by weighted summation) to form the final output feature map, which has optimized frequency components and is more suitable for complex spatial processing.

[0115] Step 2-3-2 Build the adaptive selection kernel module;

[0116] The input of this module is the output F3 of the multi-core context-aware network, Global Pool represents the global pooling operation, Conv represents the convolution operation, ReLU is the nonlinear activation function, and σ is the Sigmoid function, which is used to compress the output of the convolution operation to between 0 and 1; as shown in the following formula:

[0117] G=σ(Conv(ReLU(Conv(lobal Pool(F3))))) (4)

[0119] The adaptive selection kernel module uses global pooling to obtain the global context information of the input feature map, which is further processed through a series of convolution and ReLU layers, and then a regulation factor G is generated through the Sigmoid activation function.

[0120] The initial convolution kernel K is decomposed into low-frequency Kl and high frequency K h Two parts. The combination of these two kernels is dynamically adjusted according to the global feature map adjustment factor G, as shown in formula (5).

[0121] K′=G·K l +·K(1-G)·K h (5)

[0122] K l =Average(K) (6)

[0123] K h =KK l (7)

[0124] Among them, K′ is the final adaptive convolution kernel used to process the input features. G is the adjustment factor extracted from the global features to control the influence of the low-frequency kernel. 1-G naturally controls the influence of the high-frequency kernel to ensure the balance of low-frequency and high-frequency components.

[0125] The core of this module is to use the global feature map G to balance the low-frequency K l and high frequency K h The contribution of the kernel. The value of G determines the relative importance of low-frequency information to high-frequency information in a specific image region. For example, if a region has a higher G value, it means that the region is more inclined to low-frequency information, which is usually used for flat or uniform areas in the image; on the contrary, regions with lower G values ​​will enhance the use of high-frequency kernels and are suitable for capturing areas with more edges and details.

[0126] Step 2-3-3 Build the adaptive selection kernel module

[0127] The Adaptive Dilation Rate Selection module dynamically adjusts the dilation rate of the dilated convolution based on the local information of the input features to optimize the network's response to different spatial scales, especially when processing image regions with different textures and detail densities. This enables the network to optimally adjust its receptive field in different regions of the image, thereby enhancing the network's spatial adaptability. This adjustment is tailored so that high-frequency regions use smaller dilation rates to preserve details, while low-frequency regions use larger dilation rates to expand the receptive field.

[0128] The whole process of adaptively selecting the expansion rate module is shown in the following formula (8):

[0129] Y(p)=∑ k∈K X(p+r(p)·K′)·W(K′) (8)

[0130] Among them, Y(p) is the value of the feature map output by the adaptive dilation rate module at position p. K′ is the adaptive convolution kernel output by the adaptive kernel module. X(p) is the value of the input feature map at position p. W(K′) is the weight of the convolution kernel at position k. r(p) is the dilation rate of the dilation map at position p, which is dynamically generated by the previous convolution + ReLU layer. Y(p) is the feature map finally output by the spectrum adaptive dilation structure.

[0131] Step 2-4 constructs spatial pyramid fast pooling, progressive asymmetric feature pyramid network, and decoupling detection head; the structure of spatial pyramid fast pooling, progressive asymmetric feature pyramid network, decoupling detection head and decoupling head uses the structure in the document "A COMPREHENSIVE REVIEW OF YOLO: FROM YOLOV1 TO YOLOV8 AND BEYOND". Preferably, the input of spatial pyramid fast pooling is F4 output by the spectrum adaptive expansion structure with 256 channels, and the output is F5 with 256 channels. The input of progressive asymmetric feature pyramid network comes from feature map F2 output by multi-core context perception structure 2, F3 output by multi-core context perception, and output F5 of spatial pyramid fast pooling, with channel numbers of 64, 128, and 256, respectively. The three feature maps of 20*20, 40*40, and 80*80 output by progressive asymmetric feature pyramid network are used as inputs of three decoupling detection heads.

[0132] Step 2-5 Decoupling detection

[0133] The detection head uses the three feature maps of 20*20, 40*40, and 80*80 output by the progressive asymmetric feature pyramid network as the input of the three decoupled detection heads. Traditional detection heads often use the same set of features to complete the classification and regression tasks of the target at the same time. However, due to the different requirements of the classification and regression tasks, this way of sharing features may cause conflicts between the two, which in turn affects the detection accuracy. The decoupled detection head separates the classification and regression tasks and uses different branches to handle the two tasks respectively, thereby improving the detection performance. The classification branch focuses on the classification task of the target, inputs the feature map output by the feature pyramid network, and extracts the features for classification through a series of convolutional layers. The output features are used to predict the target category score in each detection box. The output shape of the classification branch is a four-dimensional vector of (batch size, number of anchors * number of categories, width of feature map, height of feature map). The regression branch focuses on the regression task of the bounding box, that is, predicting the location of the target. Similarly, the feature map output by the feature pyramid network is input, and the features for regression are extracted through a series of different convolutional layers. The features output by the regression branch are used to predict the location information of each bounding box (the coordinates of the upper left and lower right corners or the center point and width and height of the bounding box). The output shape is usually (batch size, number of anchors * 4, width of the feature map, height of the feature map).

[0134] Steps 2-6 Post-processing

[0135] After the decoupled detection head outputs the detection results, it is usually necessary to go through post-processing steps to obtain the final target detection results. The key steps of post-processing include: bounding box decoding, non-maximum suppression, threshold filtering, and result output, and the four steps are performed in series. The bounding box decoding operation requires decoding the position information of the detection box predicted by the regression branch into actual coordinates relative to the original image. For example, the regression head outputs the center point offset and aspect ratio, which need to be converted into the actual coordinates of the upper left corner and lower right corner. Then the non-maximum suppression operation is performed. First, the bounding box with the highest score is selected as the current box, and then other boxes with an overlap (IoU) of more than 0.5 are removed. This process is repeated until there are no candidate boxes that meet the conditions. Finally, the bounding boxes with a confidence threshold lower than 0.001 are eliminated, and the retained boxes are the detection results.

[0136] Step 3 loss function;

[0137] The loss function of the present invention includes category classification loss and bounding box regression loss:

[0138] The classification loss uses cross entropy loss to measure the model's prediction accuracy for the target category. The formula is:

[0139]

[0140] Where N is the number of bounding boxes, y i is the true label (1 means the target exists. 0 means the target does not exist), p i is the confidence score of the model's prediction.

[0141] (2) Bounding box regression loss

[0142] The bounding box regression loss is in the form of Generalized Intersection over Union (CIOU Loss) and Distance-based Feature Learning Loss (DFL Loss).

[0143] DFL Loss is used to optimize the bounding box regression task. It does not directly predict the coordinates of the bounding box, but predicts the distribution of the coordinates (for example, through a set of basis functions), so that the position of the bounding box can be more accurately estimated through the combination of these basis functions. The formula of DFL is shown in formula (10): DFL(S i ,S i+1 )=-(y i+1 -y)log(S i )+(yy i )log(S i+1 ) (10)

[0144] Here, S i and S i+1 Represent the probability of the predicted distribution in a certain interval, and y is the actual label or value. This loss function improves the prediction accuracy of the model by optimizing the distance between the predicted distribution and the true value.

[0145] CIoU Loss is a loss function for bounding box regression that takes into account the overlapping area, center point distance and aspect ratio between bounding boxes, thus providing a more comprehensive loss evaluation, as shown in formula (11):

[0146]

[0147] Among them, ρ(b,b gt ) represents the distance between the predicted box and the true box, c is the diagonal length of the smallest closed box containing the two boxes, and v and α are terms related to the aspect ratio to adjust the contribution of the aspect ratio contrast loss.

[0148] The overall loss function is shown in formula (12):

[0149]

[0150] Step 4: Use loss function The network model is trained using the training set, and the validation set is used during the training process to observe the model performance and obtain the best performing network model. Finally, the test set is used to test the effectiveness of the model. Specific embodiment:

[0152] Dataset preparation:

[0153] The public dataset HRSID is used for training and testing the network structure of the present invention. HRSID is designed for ship detection, semantic segmentation, and instance segmentation tasks using high-resolution synthetic aperture radar (SAR) images. This dataset is widely used in the development and benchmarking of SAR image analysis algorithms, especially for detecting ships. The dataset contains 5,604 high-resolution SAR images and annotations of 16,951 ship instances. These images were acquired from Sentinel-1B, TerraSAR-X, and Tandemi-X sensors, ensuring a variety of imaging conditions and resolutions. The HRSID dataset is known for its high-resolution images and the different scenes it captures (including different sea conditions, sea areas, and coastal ports), of which 3,642 are training sets and 1,962 are test sets.

[0154] At this point, the dataset preparation step is completed.

[0155] Data Augmentation

[0156] Before training, the input image needs to be enhanced. The specific steps are:

[0157] First, the images are resized to a size suitable for model input (640×640) and normalized to facilitate model learning. Next, the images are visually enhanced by adding noise, adjusting contrast, and applying sharpening and blurring effects. In addition, to simulate actual imaging conditions, the images are randomly rotated, flipped, and scaled to further increase data diversity. Random adjustments to the color space and simulated compression are also part of the enhancement step to accommodate different imaging and transmission conditions. Finally, the images are mixed using the MixUp technique to further enrich the dataset. The entire enhancement process is monitored and adjusted through logging to ensure that the enhancement operations are effective and meet expectations, thereby providing more robust and diverse SAR image data for model training.

[0158] Construction of multi-core SAR target detection network based on adaptive convolution

[0159] The network structure diagram is as follows Figure 2As shown in the figure, the input SAR image (640×640) is first received and fed into a feature extraction network, which contains multiple convolutional layers and batch normalization layers. These layers are followed by SiLU activation functions to introduce nonlinearity and enhance the expressiveness of the model. The core network, namely the multi-core context-aware network, uses convolution kernels of different sizes for parallel processing to capture multi-scale target features, and uses the context center anchor attention module to capture long-range context information, thereby improving the accuracy of detection. In addition, the method adopts a spectrum adaptive expansion structure, which dynamically adjusts the expansion rate of the convolution through spectrum analysis to optimize the balance between the receptive field and the effective bandwidth. The spatial pyramid fast pooling (SPP) technology further enhances the model's perception of spatial information. The progressive asymmetric feature pyramid network is used to construct the feature pyramid, which improves the detection performance of targets of different sizes through multi-scale fusion. The detection head is designed to be decoupled, allowing it to be optimized independently of the feature extraction network. Finally, the framework is able to output accurate target detection results, including detailed information such as the location, category, and direction of the target.

[0160] Network training

[0161] In the process of model training for implementing the multi-core SAR target detection method based on adaptive convolution, the parameters of model training include: setting the number of iterations to 300 rounds, and setting the parameters of 50 rounds to implement the early stopping strategy. Each batch contains 64 images, and the input image size is set to 640×640 pixels. During the training process, the best checkpoints and prediction results of the model are selected for saving without setting a specific saving period. The training runs on the specified device 4-card GeForce RTX 3090. In the experimental settings, SGD is selected as the optimizer. Finally, the initial learning rate is set to 0.01, the learning rate decay factor is 0.01, the momentum is 0.937, the weight decay is 0.0005, the number of warm-up rounds is 3, the warm-up momentum is 0.8, the warm-up bias learning rate is 0.1, and other loss weight parameters specific to target detection are set.

[0162] (5) Network testing

[0163] 5.1: Input image: SAR image (640×640)

[0164] 5.2: Load the best network parameters saved during the training phase or the network parameters saved in the last epoch, and test the input image.

[0165] 5.3: Network output results: such as Figure 8 As shown in the second column, the visualization of the HRSID test set detection results is shown. Figure 8The first column on the left is a comparison method (reference A COMPREHENSIVE REVIEW OF YOLO: FROM YOLOV1TO YOLOV8 AND BEYOND). It can be seen from the visualization results that the method of this embodiment is superior to the comparison method in SAR images.

Claims

1. A multi-core SAR target detection method based on adaptive convolution, characterized in that The steps include: Step 1: Dataset preparation; The HRSID public dataset is used for training and testing of the network structure; Step 2: Construction of multi-core SAR target detection model based on adaptive convolution; The target detection model includes convolution blocks, multi-core context-aware networks, spectrum adaptive expansion structures, progressive asymmetric feature pyramid networks, and decoupled detection heads. After the SAR image is input, it passes through the convolution blocks and then is input into the multi-core context-aware network. The output of the multi-core context-aware network is sequentially processed through the spectrum adaptive expansion structure and spatial pyramid fast pooling. The image that has passed through the multi-core context-aware network and spatial pyramid fast pooling is input into the progressive asymmetric feature pyramid network for processing to obtain a feature map, which is then input into the decoupled detection head for post-processing. Step 3: Loss function; Step 4: Use loss function The network model is trained using the training set and the validation set is used during the training process to observe the model performance and obtain the network model with the highest map50 and map50:95 indicators.

2. The multi-core SAR target detection method based on adaptive convolution according to claim 1, characterized in that: In step 2, the construction process of the target detection model is as follows: Step 2-1: Construct a convolution block, and connect three convolution blocks in series; The convolution block includes convolution layer, batch normalization and Silu activation function. The convolution layer window size is 3×3, the padding is 1, and the stride is 1. Step 2-2: Build a multi-core context-aware network; The multi-core context-aware network is composed of three multi-core context-aware structures connected in series; Step 2-3: Construct a spectrum adaptive expansion structure; The spectrum adaptive dilation structure is designed to process highly variable spatial frequency information and is divided into three functional blocks: an adaptive dilation rate module, a frequency selection module, and an adaptive kernel module. The frequency selection module analyzes and optimizes the performance of the input feature map at different frequencies, adjusts the contribution of different frequency bands by adjusting the weights of the frequency components, and outputs a weighted synthetic feature map. The adaptive kernel module adjusts the components of the convolution kernel to better capture high-frequency and low-frequency information, and outputs an adaptive convolution kernel. The feature map output by the frequency selection module and the adaptive convolution kernel output by the adaptive kernel module are inputs to the adaptive dilation rate module. First, an expansion map is generated according to the local characteristics of the input feature map. Each value in the expansion map defines the expansion rate that should be used for the dilated convolution at the corresponding position. According to the expansion rate indicated by the expansion map, the dilated convolution is applied to the input feature map. The dilated convolution increases its receptive field by inserting spaces between the elements of the convolution kernel without increasing the number of parameters or computational complexity. The final output of the adaptive dilation structure is a feature map that has undergone adaptive dilated convolution. Step 2-4: Construct spatial pyramid fast pooling, progressive asymmetric feature pyramid network and decoupled detection head; The input of the spatial pyramid fast pooling is the F4 output of the spectrum adaptive expansion structure with 256 channels, and the output channel number is 256. The input of the progressive asymmetric feature pyramid network comes from the feature map F2 output by the multi-core context-aware structure 2, the F3 output by the multi-core context-aware structure 2, and the output F5 of the spatial pyramid fast pooling, with the channel numbers of 64, 128, and 256 respectively. The three feature maps of sizes 20*20, 40*40, and 80*80 output by the progressive asymmetric feature pyramid network are used as the input of the three decoupled detection heads. Step 2-5: Decoupling detection; The detection head uses the three feature maps of 20*20, 40*40, and 80*80 output by the progressive asymmetric feature pyramid network as the input of three decoupled detection heads. The decoupled detection head separates the classification and regression tasks and uses different branches to process the two tasks respectively, thereby improving the detection performance. The classification branch focuses on the task of classifying the target. It inputs the feature map output by the feature pyramid network, extracts the features used for classification through a series of convolutional layers, and the output features are used to predict the target category score in each detection frame. The output shape of the classification branch is a four-dimensional vector, which includes the batch size, the number of anchors * the number of categories, the width of the feature map, and the height of the feature map. The regression branch focuses on the regression task of the bounding box, that is, predicting the location of the target. Similarly, it inputs the feature map output by the progressive asymmetric feature pyramid network, extracts the features used for regression through a series of different convolutional layers, and the features output by the regression branch are used to predict the location information of each bounding box. The location information includes the coordinates of the upper left corner and lower right corner of the bounding box, or the center point and width and height. The output shape is the batch size, the number of anchors * 4, the width of the feature map, and the height of the feature map. Step 2-6: Post-processing; After the decoupled detection head outputs the detection results, it needs to go through a post-processing step, which includes bounding box decoding, non-maximum suppression, threshold filtering and result output. The four steps are performed in series. Bounding box decoding decodes the position information of the detection box predicted by the regression branch into actual coordinates relative to the original image; then a non-maximum suppression operation is performed. First, the bounding box with the highest score is selected as the current box, and then other boxes with an overlap of more than 0.5 with it are removed, and the non-maximum suppression process is repeated in sequence until there are no candidate boxes that meet the conditions; finally, the bounding boxes with a confidence threshold lower than 0.001 are eliminated, and the retained boxes are the detection results.

3. The multi-core SAR target detection method based on adaptive convolution according to claim 1, characterized in that: The loss function described in step 3 includes category classification loss and bounding box regression loss: The classification loss uses cross entropy loss to measure the model's prediction accuracy for the target category: Where N is the number of bounding boxes, y i is the true label, 1 means the target exists, 0 means the target does not exist, p i is the confidence score of the model prediction; The bounding box regression loss is in the form of a generalized intersection loss over a union and a distance-based feature learning loss; DFL Loss is used to optimize the bounding box regression task. It does not directly predict the coordinates of the bounding box, but predicts the distribution of the coordinates, thereby more accurately estimating the position of the bounding box through a combination of basis functions. The formula of FL is: DFL(S i ,S i+1 )=-(y i+1 -y)log(S i )+(y-y i )log(S i+1 ) Here, S i and S i+1 Represent the probability of the predicted distribution in the i-th interval and the i+1-th interval respectively, and y is the true label or value; the loss function improves the prediction accuracy of the model by optimizing the distance between the predicted distribution and the true value; CIoU Loss is a loss function for bounding box regression that takes into account the overlap area, center point distance, and aspect ratio between bounding boxes, thus providing a more comprehensive loss assessment: Among them, ρ(b,b gt ) represents the predicted box b gt The distance between the real box b, c is the diagonal length of the smallest closed box containing the two boxes, v and α are terms related to the aspect ratio, which are used to adjust the contribution of the aspect ratio contrast loss; The overall loss function As shown in the formula:

4. The multi-core SAR target detection method based on adaptive convolution according to claim 3, characterized in that: In step 2-2, in the multi-core context-aware structure, the input is the feature map F from the previous stage l-1 , l indicates that the current layer is the lth layer of the multi-core context-aware structure. The input feature map is first reduced in spatial size by the Haar wavelet downsampling module; the 3×3 convolution layer performs a convolution operation on the feature map after the Haar wavelet downsampling to extract semantic features, where the 3×3 convolution layer changes the depth of the feature map, that is, the number of channels; the feature map X after the 3×3 convolution layer is l-1 Divide into two parts evenly according to the number of channels and The two parts are then assigned to two different processing paths; one path is a feed-forward network that processes the input feature map through two convolutional layers and activation functions in series. Perform direct forward propagation to extract and transfer feature information, and the output is The other path consists of multiple multi-core context-aware layers connected in series, each of which is a basic building block in the network, and the output is The output of the feed-forward network FFN Output of the multi-core context-aware layer Then, the convolutional layer is used to further integrate the features and output the final feature map F. l , the output F of the entire multi-core context-aware structure l The expression is shown in the formula: Among them, Concat represents a concatenation operation, and Conv represents a convolution operation. The concatenation and convolution operations help to integrate the features of different processing paths and enhance the representation capability of the model of the present invention for input data.

5. The multi-core SAR target detection method based on adaptive convolution according to claim 4, characterized in that: The multi-core context perception layer integrates the multi-core context perception module and the context center anchor attention module. The two modules are executed in parallel. The input of the multi-core context perception layer is the feature map passed in from the previous structure or the previous multi-core context perception layer. The multi-core context-aware module uses multi-scale convolution kernels to capture target features of different scales; The input of the context center anchor attention module is also the feature map passed in by the previous structure or the previous multi-core context perception layer. Focus on capturing the contextual relationship between distant pixels in the image; the context center anchor attention module first obtains global features through average pooling, and then extracts and refines global features through 1×1 convolution layer and 3×3 depth-separable convolution, thereby enhancing the semantic information of the features; multiply the feature maps output by the multi-core context perception module and the context center anchor attention module, and then add the multiplication result to the feature map output by the multi-core context perception module, and process the fused feature map through a 1x1 convolution layer to further integrate the features and reduce the feature dimension, and output The processing results of the two parallel paths are finally combined through channel concatenation and fusion operations to generate the output feature maps of each stage.

6. The multi-core SAR target detection method based on adaptive convolution according to claim 5, characterized in that: In the multi-core context-aware module, the input of the multi-core context-aware module is The feature map output after processing by the multi-core context-aware module is P l-1,n , used as input for subsequent layers or modules; the initial depth-separable convolution operation uses a 3x3 convolution kernel to align the input Feature extraction is performed, and then multi-scale depthwise separable convolutions are performed using 5x5, 7x7, 9x9 and 11x11 convolution kernels. These operations help the model capture spatial features of different scales. Identity mapping is used to directly pass the 3×3 depthwise separable convolution to the fusion stage and fuse it with feature maps of other scales.

7. The multi-core SAR target detection method based on adaptive convolution according to claim 5, characterized in that: In the context center anchor attention module, the formula of the context center anchor attention module is expressed as: Average pooling operation on the input feature map Downsampling of the spatial dimension is performed to reduce the size of the feature map to reduce the amount of subsequent calculations. Average pooling is achieved by calculating the average value of each small area in the input feature map, and then a 1x1 convolution layer is used to adjust the number of channels. The first is a 1x(11+2xn) depthwise separable convolution to process the features in the width direction; followed by a (11+2xn)x1 depthwise separable convolution to process the features in the height direction. The depthwise separable convolution performs spatial convolution on each input channel separately, and then applies a 1x1 convolution to fuse the channel information. The features after the depthwise convolution are integrated again through a 1x1 convolution, which further improves the expressiveness of the features and the ability to integrate features of different sizes. Finally, the final convolution output is processed through the Sigmoid activation function to obtain the final attention weight map A. l-1,n , the Sigmoid function compresses the output value between 0 and 1, indicating the importance of each position.

8. The multi-core SAR target detection method based on adaptive convolution according to claim 2, characterized in that: In step 2.3, the steps of constructing the spectrum adaptive expansion structure are: Step 2-3-1: Build a frequency selection module; In the frequency selection module, the input spatial domain feature map is the output F3 of the multi-core context-aware network, which is first converted to the frequency domain through Fourier transform to form X F , the transformation reveals the different frequency components in the image; then, X F is decomposed into a number of predefined frequency bands, each of which is processed by applying a corresponding frequency domain mask M b Extract the features of the corresponding frequency range Each frequency band feature is independently passed through the convolution kernel W designed for it b Processing is performed to generate the convolution result Y b To determine the contribution of each frequency band to the final output, b Apply 1×1 convolution and Sigmoid activation function to generate selection map S b , select Figure S b Expresses the importance of each frequency band; ultimately, all weighted frequency band features Y b By comparing their respective selections to b Multiply and accumulate to synthesize the final output feature map Y, so that the module can dynamically adjust the focus on different frequency bands and optimize the overall feature extraction and information synthesis. The final output feature map is calculated by formula (3): And=∑ b S b ·AND b (3) Among them, S b =Sigmoid(Conv(X b )) represents the selection weight obtained by applying 1x1 convolution and Sigmoid function to each frequency band feature. The re-weighted frequency band features are weighted summed to form the final output feature map, which has optimized frequency components and is more suitable for complex spatial processing; Step 2-3-2: Build the adaptive selection kernel module; The input of the adaptive selection kernel module is the output F3 of the multi-core context-aware network, Global Pool represents the global pooling operation, Conv represents the convolution operation, ReLU is the nonlinear activation function, and σ is the Sigmoid function, which is used to compress the output of the convolution operation to between 0 and 1; as shown in formula (4): G=σ(Conv(ReLU(Conv(local Pool(F3))))) (4) The adaptive selection kernel module uses global pooling to obtain the global context information of the input feature map, which is further processed through a series of convolution and ReLU layers, and then a regulation factor G is generated through the Sigmoid activation function; The initial convolution kernel K is decomposed into low-frequency K l and high frequency K h The two parts dynamically adjust the combination of the two kernels according to the global feature map adjustment factor G, as shown in formula (5): K′=G·K l +·K(1-G)·K h (5) K l =Average(K) (6) K h =K-K l (7) Among them, K ′ It is the final adaptive convolution kernel, which is used to process the input features. G is the adjustment factor extracted from the global features, which controls the influence of the low-frequency kernel. 1-G naturally controls the influence of the high-frequency kernel to ensure the balance between low-frequency and high-frequency components. Average is the mean value. The core of the adaptive kernel selection module is how to use the global feature map to balance the low-frequency K l and high frequency K h The kernel contribution,G,value determines the relative importance of low-frequency information and high-frequency information,in a specific image region; Step 2-3-3: Construct an adaptive expansion rate selection module; The whole process of adaptively selecting the expansion rate module is shown in formula (8): Y(p)=∑ k∈K X(p+r(p)·K′)·W(K′) (8) Among them, Y(p) is the value of the feature map output by the adaptive expansion rate module at position p, K ′ is the adaptive convolution kernel output by the adaptive selection kernel module, X(p) is the value of the input feature map at position p, and W(K ′ ) is the weight of the convolution kernel at position k, r(p) is the dilation rate of the dilation map at position p, which is dynamically generated by the previous convolution + ReLU layer, and Y(p) is the feature map finally output by the spectrum adaptive dilation structure.

9. An electronic device, characterized in that: include: one or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, and the program codes can be called by a processor to execute the method according to any one of claims 1 to 8.

Images