Sar image target detection model, method, electronic device and storage medium

Abstract

The application discloses a SAR image target detection model and method, an electronic device and a storage medium, belongs to the field of SAR image target detection, takes YOLO11 as a baseline network, YOLO11 includes a backbone feature extraction network, a neck feature fusion network and a detection head, and the backbone feature extraction network is formed by alternately stacking LSConv and Rep_C3k2. The above-mentioned SAR image target detection model, method, electronic device and storage medium balance the precision, inference speed and anti-ground object interference and speckle noise capability of SAR image target detection through the three-order architecture design of the LSConv+Rep_C3k2 lightweight backbone network, the MHMC_C3k2 multi-scale feature fusion neck and the ELA attention detection head.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of SAR image target detection, and in particular to a SAR image target detection model, a method, an electronic device and a storage medium. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) image target detection has key application value in the fields of military reconnaissance, ocean monitoring, resource exploration, etc., but the existing technology still faces the following problems to be solved urgently:

[0003] Contradiction between network complexity and actual application requirements: In order to improve the detection accuracy, the existing SAR image target detection network generally adopts a complex deep architecture or a redundant feature extraction module. Although this kind of complex network can optimize the detection accuracy to a certain extent, it brings significant calculation and memory overheads - the inference process is time-consuming and difficult to meet the real-time requirement of high real-time scenes (such as battlefield rapid reconnaissance, unmanned aerial vehicle real-time monitoring); at the same time, the super-high demand for memory resources not only depends on high-performance computing devices to support, but also leads to a sharp drop in detection speed and unstable effect when deployed on memory-limited platforms such as mobile devices and embedded systems, which seriously limits the universality of the technology. SUMMARY

[0004] The purpose of the present application is to provide a SAR image target detection model, method, electronic device and storage medium to solve the above technical problems.

[0005] To achieve the above purpose, the present application provides a SAR image target detection model, taking YOLO11 as a baseline network, YOLO11 including a backbone feature extraction network, a neck feature fusion network and a detection head, the backbone feature extraction network being alternately stacked by LSConv and Rep_C3k2, wherein the LSConv is composed of a large kernel perception sub-module and a small kernel aggregation sub-module, and the large kernel perception sub-module first reduces the channel number of the feature map to C / 2 through 1x1 pointwise convolution, then applies 7x7 large kernel depth separable convolution to extract context information on the perception area, and finally generates context adaptive weights for small kernel aggregation through 1x1 convolution; the small kernel aggregation sub-module divides the channels of the input feature map into G groups, each group of C / G channels sharing aggregation weights, dynamically aggregates the features within the small kernel neighborhood using the context adaptive weights generated by the large kernel perception sub-module, and then uses 3x3 small kernel dynamic convolution for feature aggregation, and finally outputs the aggregated feature map.

[0006] The method of the SAR image target detection model includes the following steps:

[0007] S1, obtaining a SAR image dataset and performing preprocessing;

[0008] S2, input the preprocessed SAR image into a backbone feature extraction network, perform light feature extraction on the preprocessed SAR image by using LSConv of the backbone feature extraction network, and perform feature fusion by using Rep_C3k2 to obtain a multi-scale feature set;

[0009] S3, performing cross-layer feature integration on the multi-scale feature set by using a neck feature fusion network, and enhancing target features by a bidirectional feature pyramid structure and an MHMC_C3k2 submodule to output a three-scale optimized feature map;

[0010] S4, inputting the three-scale optimized feature map into a corresponding scale ELA_Detect detection head, suppressing speckle noise through an attention mechanism, and outputting a target detection result.

[0011] An electronic device comprising a processor and a memory, wherein the processor is configured to perform the method of the SAR image target detection model.

[0012] A storage medium having computer program instructions stored thereon, the computer program instructions being executed by a processor to implement the method of the SAR image target detection model.

[0013] Therefore, the SAR image target detection model, method, electronic device and storage medium have the beneficial effects that:

[0014] 1. Realize a light network structure, improve detection efficiency and deployment adaptability: through the structure reparameterization design of LSConv combined with Rep_C3k2 “training multi-branch to preserve accuracy, and single-branch to speed up inference”, the model parameter quantity and calculation quantity can be effectively reduced; at the same time, the dependence on memory is reduced, which can adapt to mobile devices, embedded systems and other scenes with limited memory resources, and the inference speed is significantly improved, meeting the real-time detection demand;

[0015] 2. Multi-scale feature fusion enhancement reduces false detection and missed detection caused by ground objects: the MHMC_C3k2 submodule realizes the deep fusion of high-level semantic features (distinguish targets and ground objects) and low-level detail features (accurately locate targets) by “3x3 / 5x5 / 7x7 multi-scale convolution kernel cross-scale feature extraction + channel splicing aggregation”, combined with “top-down-bottom-up” bidirectional feature pyramid; effectively weaken the confusion of building, tree and other ground object background and target features, significantly reduce the false detection rate and the missed detection rate, and improve the reliability of SAR image target detection;

[0016] 3. Inhibit speckle noise and improve small target detection accuracy: The ELA attention mechanism integrated in the detection head captures long-range spatial dependencies through "horizontal / vertical strip pooling + attention map weighting", accurately suppresses speckle noise generated by SAR image coherent imaging, and focuses on key regional features such as target boundary and texture. Especially for small targets (easily covered by noise), it can effectively improve the feature perception ability and further ensure the accuracy of the detection result.

[0017] 4. Strong network architecture compatibility and generalization ability: Each module (LSConv, Rep_C3k2, MHMC_C3k2, ELA) is designed as "plug and play", which can be flexibly adapted to different SAR image scenes (such as different ground object density and different target size distribution scenes) based on YOLO11 baseline. At the same time, through data enhancement and multi-scale detection head design, the adaptability of the model to the distribution difference of SAR image data is enhanced, and the generalization ability is stronger.

[0018] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The framework diagram of the SAR image target detection model of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be further described in detail below with the help of the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present application and not to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application. The examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout.

[0021] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0022] The embodiments of the present application will be described in detail below with the help of the accompanying drawings.

[0023] As Figure 1As shown, the SAR image target detection model takes YOLO11 as the baseline network, YOLO11 includes a backbone feature extraction network, a neck feature fusion network and a detection head, the backbone feature extraction network is alternately stacked by LSConv (Lightweight Separable Convolution, function: reduce the model parameter quantity and calculation overhead on the premise of ensuring the SAR image target feature extraction capability, provide "lightweight, high information density" multi-scale basic features for subsequent feature fusion, adapt to the real-time deployment needs of resource-constrained devices (such as embedded systems)) and Rep_C3k2 (Reparameterized C3k2 Module, function: fuse the shallow detail features and deep semantic features of the SAR image, and at the same time, through the structure reparameterization technology, ensure the feature extraction accuracy in the training stage, eliminate the multi-branch calculation overhead in the inference stage, and improve the model inference speed), wherein the LSConv is composed of a large kernel perception sub-module and a small kernel aggregation sub-module, and the large kernel perception sub-module first reduces the channel number of the input SAR feature map to C / 2 through 1×1 pointwise convolution, then applies 7×7 large kernel depth separable convolution to extract the context information on the perception area, and finally generates the context adaptive weight for small kernel aggregation through 1×1 convolution; that is, the large kernel perception sub-module is composed of 1×1 pointwise convolution, 7×7 large kernel depth separable convolution, and 1×1 convolution combined with batch normalization (BN) and ReLU activation function in series, function: expand the receptive field through 7×7 large kernel, extract the context information of the target in the SAR image; at the same time, realize channel dimension compression (reduce the input channel number to 1 / 2) and nonlinear feature enhancement through 1×1 convolution; the small kernel aggregation sub-module divides the channels of the input feature map into G groups, each group of C / G channels shares the aggregation weight, uses the context adaptive weight generated by the large kernel perception sub-module to dynamically aggregate the features in the small kernel neighborhood, and then uses 3×3 small kernel dynamic convolution to aggregate the features, and finally outputs the aggregated feature map; that is, the small kernel aggregation sub-module is composed of channel grouping, dynamic weighted aggregation, and 3×3 small kernel dynamic convolution in series, which is used to group the input features by channels to reduce the calculation cost, dynamically weight the small kernel neighborhood features using the weight generated by the large kernel perception sub-module, and finally complete feature aggregation through 3×3 small kernel convolution to output lightweight feature maps with high information density.

[0024] The Rep_C3k2 comprises a channel pre-aggregation sub-module, a feature segmentation sub-module, a multi-branch processing sub-module, and a feature splicing and fusion sub-module; the channel pre-aggregation sub-module adopts a 1x1 convolution layer to preliminarily aggregate channel information of an aggregated feature map output by the LSConv, and compresses the channel dimension to reduce subsequent calculation amount; the feature segmentation sub-module is used to divide the feature map output by the channel pre-aggregation sub-module into two parts A and B according to the channel number, wherein the part A is directly transmitted to the feature splicing and fusion sub-module and is used to retain shallow basic features; the part B is input into the multi-branch processing sub-module; the multi-branch processing sub-module is used to process the part B feature map by adopting the RepConv multi-branch structure, so as to extract deep layer features in the training stage; in the inference stage, the multi-branch is equivalent to a single-branch structure through structure reparameterization, the additional calculation overhead brought by the multi-branch is eliminated, and a deep layer processing feature map is output; the feature splicing and fusion sub-module is used to perform second feature segmentation on the deep layer processing feature map, to obtain parts C and D, and then the parts A, C and D are spliced according to the channel dimension, so as to realize fusion of the deep layer features and the shallow layer features, and to obtain a multi-scale feature set.

[0025] The neck feature fusion network adopts a bidirectional feature pyramid structure from top to bottom and from bottom to top. The neck feature fusion network comprises a feature transmission path submodule and an MHMC_C3k2 submodule (Multi-scale Hybrid Channel C3k2 Module, function: to realize deep fusion of cross-scale and cross-channel features of the SAR image, weaken the confusion of the background of buildings, trees and other ground objects and the target features, improve the distinguishability of the target features, and reduce the false detection and missed detection rates). The feature transmission path submodule is divided into a top-down path and a bottom-up path, forming a bidirectional feature interaction link. The top-down path is used to receive a multi-scale feature set output by the backbone feature extraction network, and the resolution of high-level feature maps in the multi-scale feature set is improved through an upsampling operation. After the high-level feature maps are matched in size with middle-level feature maps in the multi-scale feature set, the high-level feature maps and the middle-level feature maps are fused through channel splicing to supplement the middle-level feature maps with “target and ground object distinguishability semantic information” and avoid misjudgment of the ground object background features as target features. The bottom-up path is used to perform dimension adjustment on the middle-level feature maps fused through the top-down path through a Conv layer, and then perform channel splicing on the bottom-level feature maps in the multi-scale feature set to supplement the high-level feature maps with “target boundary, texture and other detail information” and improve the target positioning accuracy. The MHMC_C3k2 submodule is embedded in a splicing node of the feature transmission path, realizes cross-channel and cross-scale information interaction through “branch splitting, multi-scale convolution and feature aggregation”, and comprises: a feature branch splitting subunit: used to equally divide the initially fused feature maps after splicing of the high-level feature maps and the middle-level feature maps into a and b two branches according to the number of channels, wherein the a branch is transmitted to a feature splicing and channel aggregation unit to retain local detail features in the feature maps and avoid loss of details caused by excessive convolution, and the b branch is transmitted to a multi-scale convolution processing subunit; the multi-scale convolution processing subunit: used to perform multi-scale feature extraction on the b branch to form a feature subset containing multi-scale information; a dimension rearrangement and merging unit: used to rearrange the dimensions of the feature subset containing multi-scale information, uniformly set the dimensions of features of different scales to the same format, and then perform channel merging to obtain a merged feature map integrating multi-scale information; and the feature splicing and channel aggregation unit: used to perform channel splicing of the a branch (local detail features) and the merged feature map (multi-scale features), and then perform channel aggregation through a 1x1 point-by-point convolution layer. The 1x1 convolution not only realizes reasonable compression of the channel dimension, but also enhances the nonlinear fitting capability of the network, further highlights the target features and suppresses the ground object background interference, and finally outputs a three-scale optimized feature map.

[0026] That is, the neck feature fusion network can solve the problem of cross-group channel information isolation caused by grouped convolution in the backbone feature extraction network by configuring the MHMC_C3k2 module based on the multi-head mixed convolution at the key node of feature fusion, and can enhance the three-scale optimized feature map and weaken the background interference of buildings, trees and other objects through multi-scale feature fusion.

[0027] Three parallel ELA_Detect detection heads corresponding to the scales of the three-scale optimized feature map are set, respectively, to adapt to large, medium and small scale SAR image target detection, and the core is to suppress the speckle noise generated by SAR image coherent imaging, enhance the key region features of the target (such as boundary and texture), and improve the small target detection accuracy. The ELA_Detect detection head includes an ELA attention sub-module, a feature optimization sub-module and a target prediction sub-module; the ELA attention sub-module is used for generating a spatial attention map through spatial strip pooling and local interaction for a single-scale optimized feature map (dimension HxWxC) of the three-scale optimized feature map; the feature optimization sub-module is used for element-wise multiplication of the optimized feature map output by the neck feature fusion network and the spatial position attention map generated by the ELA attention sub-module, to obtain an enhanced feature map weighted by attention-the feature map retains the core information of the target and weakens the noise interference, providing a high signal-to-noise ratio feature input for subsequent prediction; and the target prediction sub-module is used for completing the prediction of target categories and bounding boxes based on the enhanced feature map.

[0028] The ELA attention sub-module includes: a strip pooling unit used for respectively performing horizontal strip pooling and vertical strip pooling on the single-scale optimized feature map to generate horizontal 1D feature vectors and vertical 1D feature vectors; a local interaction optimization unit used for respectively performing local information interaction on the horizontal 1D feature vectors and the vertical 1D feature vectors by using a 1x1 convolution layer, the number of convolution kernels being consistent with the number of input channels C, to enhance the resolution capability of the feature vectors to the "target-noise" difference, and output optimized horizontal feature vectors and vertical feature vectors; and an attention generation unit used for respectively performing grouped normalization on the optimized horizontal feature vectors and the vertical feature vectors to stabilize the feature distribution, and then generating a horizontal attention map (1xWxC) and a vertical attention map (Hx1xC) by using a Sigmoid activation function; and then multiplying the horizontal attention map and the vertical attention map after spatial dimension expansion, to obtain a spatial position attention map of HxWxC, wherein the weight of the target key region is enhanced, and the weight of the speckle noise region is suppressed.

[0029] The method of the SAR image target detection model includes the following steps:

[0030] S1, obtaining a SAR image data set and performing preprocessing;

[0031] S2, input the preprocessed SAR image into a backbone feature extraction network, perform light feature extraction on the preprocessed SAR image by using LSConv of the backbone feature extraction network, and perform feature fusion by using Rep_C3k2 to obtain a multi-scale feature set;

[0032] Step S2 specifically includes the following steps:

[0033] S21, the preprocessed SAR image is first subjected to an initial Conv module (3x3 convolution) to complete preliminary feature mapping and dimension adjustment, and an initial feature map is generated;

[0034] S22, the initial feature map is input into the first LSConv to perform light feature extraction;

[0035] S221, macro kernel perception sub-module processing: for the initial feature map HxWxC, first reduce the channel number to C / 2 through 1x1 point convolution to reduce the calculation amount; then apply 7x7 macro kernel depth separable convolution to extract context information in the perception area and expand the receptive field; finally, generate context adaptive weights through 1x1 convolution, and simultaneously normalize and activate the features through the BN layer to enhance the nonlinear expression of the features;

[0036] S222, small kernel aggregation sub-module processing: the channels of the feature map input into the small kernel aggregation sub-module are divided into G groups, and C / G channels in each group share aggregation weights; the context adaptive weights generated in step S221 are used to dynamically aggregate the features in the small kernel neighborhood; then, feature aggregation is performed through 3x3 small kernel dynamic convolution, the calculation cost is further reduced through the grouping mechanism, the response capability of the model to key areas is enhanced, and an aggregated feature map is output;

[0037] S23, input the aggregated feature map output by the first LSConv into the first Rep_C3k2 to perform feature deepening and fusion;

[0038] S231, the channel pre-aggregation sub-module aggregates and compresses the channel information of the aggregated feature map through 1x1 convolution to reduce the subsequent calculation amount;

[0039] S232, the feature segmentation sub-module equally divides the feature map output by the channel pre-aggregation sub-module into two parts A and B according to the channel number, the A part retains the shallow basic features and transmits them to the feature splicing and fusion sub-module; and the B part is input into the multi-branch processing sub-module;

[0040] S233, the multi-branch processing sub-module processes the B part using the RepConv multi-branch structure: in the training stage, deep features are extracted through parallel paths; in the inference stage, the multi-branch structure is equivalent to a single-branch structure through structure reparameterization, and a deep processing feature map is output;

[0041] S234, the feature splicing and fusion sub-module performs second cutting on the deep processing feature map to obtain a C part and a D part, and then splices the A part, the C part and the D part in the channel dimension to output a fused feature map;

[0042] S24, steps S22 and S23 are repeatedly performed to generate fused feature maps of different scales, and finally a multi-scale feature set is output.

[0043] S3, the neck feature fusion network is used to perform cross-layer feature integration on the multi-scale feature set, and the target feature is enhanced through the bidirectional feature pyramid structure and the MHMC_C3k2 sub-module, and three-scale optimized feature maps are output.

[0044] Step S3 specifically includes the following steps:

[0045] S31, the multi-scale feature maps in the multi-scale feature set are divided into high-level feature maps, middle-level feature maps and bottom-level feature maps according to resolution and semantic information;

[0046] S32, top-down path fusion: processing the high-level feature maps in the multi-scale feature set to supplement the semantic distinguishing ability of the middle-level feature maps;

[0047] S321, up-sampling adaptation: performing up-sampling operation on the high-level feature maps to improve their resolution to the resolution of the middle-level feature Figure 1 ;

[0048] S322, channel splicing: splicing the up-sampled high-level feature maps and the middle-level feature maps in the channel dimension to obtain a preliminary fused feature map;

[0049] S323, inputting the preliminary fused feature map into the MHMC_C3k2 sub-module to perform multi-scale feature enhancement and ground object interference suppression;

[0050] S3231, feature branch cutting: through the feature branch cutting subunit, the preliminary fused feature map is equally divided into a branch and a branch in the channel number, wherein the a branch is transmitted to the feature splicing and channel aggregation unit, and the b branch is transmitted to the multi-scale convolution processing subunit;

[0051] S3232, multi-scale convolution extraction: through the multi-scale convolution processing subunit, the b branch is first dimensionally rearranged, and then the rearranged b branch is equally divided into three small sub-branches in the channel number, and 3x3 convolution kernel (capturing small target correlation features), 5x5 convolution kernel (capturing medium target correlation features) and 7x7 convolution kernel (capturing large target and ground object background difference features) are used for convolution operation on the three small sub-branches respectively to generate three groups of feature subsets containing different scale information;

[0052] S3233, Dimension rearrangement and merging: Through the dimension rearrangement and merging unit, the 3 groups of feature subsets are dimensionally rearranged (unifying the channel dimension format of each subset), and then merged according to the channel dimension, to obtain a merged feature map integrating multi-scale information;

[0053] S3234, Feature splicing and aggregation: Through the feature splicing and channel aggregation unit, the a branch (local detail feature) and the merged feature map (multi-scale feature) are spliced according to the channel, and then a 1x1 point convolution layer is used for channel aggregation, which not only compresses the channel dimension to a reasonable range to reduce the amount of calculation, but also enhances the "target-terrain" feature discrimination degree through nonlinear transformation, and finally outputs an optimized middle-layer fusion feature map;

[0054] S33, Bottom-up path fusion: The optimized middle-layer fusion feature map output by S32 is processed to supplement the semantic information of the bottom-layer feature map (high resolution, low semantic information) in the multi-scale feature set;

[0055] S331, Dimension adjustment: A 3x3 convolution layer is used to adjust the channel dimension and resolution of the optimized middle-layer fusion feature map, so that its channel number and resolution match those of the bottom-layer feature map;

[0056] S332, Channel splicing: The optimized middle-layer fusion feature map after dimension adjustment is spliced with the bottom-layer feature map according to the channel dimension, to obtain a secondary fusion feature map;

[0057] S333, Secondary optimization of MHMC_C3k2 sub-module: Repeat step S323 to finally output an optimized small-scale fusion feature map;

[0058] S34, Separate optimization of high-layer feature map and three-scale integration: The high-layer feature map output by the backbone feature extraction network is separately optimized;

[0059] S341, Optimization of MHMC_C3k2 sub-module: The high-layer feature map is input into the MHMC_C3k2 sub-module, and S323 is repeated to output an optimized large-scale fusion feature map;

[0060] S342, Integration of three-scale optimized feature maps: The optimized middle-layer fusion feature map output by S32, the optimized small-scale fusion feature map output by S33, and the optimized large-scale fusion feature map output by S341 are integrated to form a three-scale optimized feature map.

[0061] S4, The three-scale optimized feature maps are respectively input into the corresponding scale ELA_Detect detection head, and the speckle noise is suppressed through the attention mechanism, and the target detection result is output.

[0062] It specifically includes the following steps:

[0063] S41, ELA attention enhancement: each detection head first performs horizontal / vertical strip pooling on the input corresponding to the scale optimization feature map - horizontal pooling generates a 1xWxC horizontal feature vector along the height direction, and vertical pooling generates a Hx1xC vertical feature vector along the width direction; 1x1 convolution is used for local interaction of the two types of feature vectors, the kernel size is adjusted to control the interaction range, and then grouped normalization and Sigmoid activation are performed to generate horizontal and vertical position attention maps, and the two are multiplied to obtain a spatial attention map; specifically, the horizontal strip pooling performs global average pooling on each column of features along the height direction (H dimension) to generate a 1xWxC horizontal 1D feature vector; the vertical strip pooling performs global average pooling on each row of features along the width direction (W dimension) to generate a Hx1xC vertical 1D feature vector, which focuses on the target continuous area feature through a narrow kernel strip shape, avoiding the interference of discrete areas where speckle noise is located.

[0064] S42, feature weighting and prediction: multiply the optimized feature map and the spatial attention map element by element to obtain an enhanced feature map with "noise suppression and target highlighting"; the enhanced feature map is mapped to a prediction feature map containing "class probability (matching the number of target classes to be detected), bounding box coordinate offset (center point x / y, width / height), and target confidence" through 3x3 convolution;

[0065] S43: result screening: decode the prediction feature map, convert the bounding box offset to actual coordinates in the image coordinate system, and then remove the overlapping redundant prediction boxes through non-maximum suppression (NMS, IOU threshold set to 0.5) to retain the highest confidence target prediction result;

[0066] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A SAR image target detection model, using YOLO11 as the baseline network, wherein YOLO11 includes a backbone feature extraction network, a neck feature fusion network, and a detection head, characterized in that: The backbone feature extraction network consists of alternating stacks of LSConv and Rep_C3k2. LSConv comprises a large kernel perception submodule and a small kernel aggregation submodule. The large kernel perception submodule first reduces the number of channels in the input SAR feature map to C / 2 using a 1×1 pointwise convolution, then applies a 7×7 large kernel depthwise separable convolution to extract contextual information in the perception region, and finally generates context-adaptive weights for small kernel aggregation using a 1×1 convolution. The small kernel aggregation submodule divides the channels of the input feature map into G groups, with each group containing C / G channels sharing aggregation weights. It uses the context-adaptive weights generated by the large kernel perception submodule to dynamically aggregate features within the small kernel neighborhood, and then uses a 3×3 small kernel dynamic convolution to further aggregate features, ultimately outputting an aggregated feature map. Rep_C3k2 includes a channel pre-aggregation submodule, a feature slicing submodule, a multi-branch processing submodule, and a feature splicing and fusion submodule; The channel pre-aggregation submodule uses a 1×1 convolutional layer to perform preliminary aggregation of the channel information of the aggregated feature map output by LSConv, compressing the channel dimension to reduce the amount of subsequent computation. The feature segmentation module divides the feature map output by the channel pre-aggregation module into two parts, A and B, according to the number of channels. Part A is directly passed to the feature splicing and fusion module to retain shallow basic features; part B is input to the multi-branch processing module. The multi-branch processing submodule is used to process the feature map of part B using the RepConv multi-branch structure to extract deep features during the training phase; during the inference phase, the multi-branch structure is converted into a single-branch structure through structural reparameterization, eliminating the additional computational overhead caused by the multi-branch structure and outputting the deep-processed feature map. The feature splicing and fusion submodule is used to perform a second feature segmentation on the deep processed feature map to obtain part C and part D. Then, parts A, C and D are spliced ​​together according to the channel dimension to achieve the fusion of deep features and shallow features and obtain a multi-scale feature set. The neck feature fusion network adopts a top-down and bottom-up bidirectional feature pyramid structure. The neck feature fusion network includes a feature transmission path submodule and an MHMC_C3k2 submodule. The feature transmission path submodule is divided into a top-down path and a bottom-up path, forming a bidirectional feature interaction link. The MHMC_C3k2 submodule is embedded at the concatenation node of the feature propagation path, and it includes: Feature branch slicing subunit: Used to process the initial fused feature map after splicing high-level feature maps and mid-level feature maps. It is divided into two branches, a and b, according to the number of channels. Branch a is passed to the feature splicing and channel aggregation unit, and branch b is passed to the multi-scale convolution processing subunit. Multi-scale convolution processing subunit: used to extract multi-scale features from branch b, forming a feature subset containing multi-scale information; Dimension rearrangement and merging unit: used to rearrange the dimensions of feature subsets of multi-scale information, unify the feature dimensions of different scales into the same format, and then merge the channels to obtain a merged feature map that integrates multi-scale information. Feature concatenation and channel aggregation unit: used to concatenate the a branch and the merged feature map through channels, and then perform channel aggregation through a 1×1 pointwise convolutional layer, finally outputting a three-scale optimized feature map; Three parallel ELA_Detect heads are set up for each scale of the three-scale optimized feature map. The ELA_Detect head includes an ELA attention submodule, a feature optimization submodule, and a target prediction submodule. ELA Attention Submodule: Used for single-scale optimization of feature maps for three-scale optimization feature maps, generating spatial attention maps through spatial strip pooling and local interactions; Feature optimization submodule: This module is used to perform element-wise multiplication between the optimized feature map output by the neck feature fusion network and the spatial location attention map generated by the ELA attention submodule to obtain an attention-weighted enhanced feature map. Target prediction submodule: Used to predict the target category and bounding box based on the enhanced feature map.

2. The SAR image target detection model according to claim 1, characterized in that: The top-down path receives the multi-scale feature set output by the backbone feature extraction network and improves the resolution of the high-level feature map in the multi-scale feature set through upsampling operations. After matching the size of the mid-level feature map in the multi-scale feature set, the semantic features in the high-level feature map are fused with the mid-level feature map through channel concatenation. The bottom-up path adjusts the dimensions of the mid-level feature map after fusion by the top-down path through the Conv layer, and then concatenates it with the bottom-level feature map in the multi-scale feature set through channel concatenation.

3. The SAR image target detection model according to claim 1, characterized in that: The ELA attention submodule includes: Strip pooling unit: used to perform horizontal strip pooling and vertical strip pooling on the single-scale optimized feature map to generate horizontal 1D feature vectors and vertical 1D feature vectors; Local interaction optimization unit: Used to perform local information interaction on the horizontal 1D feature vector and the vertical 1D feature vector using a 1×1 convolutional layer, and output the optimized horizontal and vertical feature vectors. Attention generation unit: It is used to group and normalize the optimized horizontal and vertical feature vectors to stabilize the feature distribution, and then generate horizontal and vertical attention maps through the Sigmoid activation function; then the horizontal and vertical attention maps are multiplied by expanding them according to the spatial dimension to obtain the spatial location attention map.

4. A SAR image target detection method, characterized in that: To execute the SAR image target detection model of claim 2, the following steps are included: S1. Obtain the SAR image dataset and perform preprocessing; S2. Input the preprocessed SAR image into the backbone feature extraction network, use the LSConv of the backbone feature extraction network to perform lightweight feature extraction on the preprocessed SAR image, and then use Rep_C3k2 to perform feature fusion to obtain a multi-scale feature set. S3. Perform cross-layer feature integration on the multi-scale feature set using the neck feature fusion network, and enhance the target features through the bidirectional feature pyramid structure and the MHMC_C3k2 submodule to output a three-scale optimized feature map. S4. Input the three-scale optimized feature maps into the corresponding scale ELA_Detect detection head, suppress speckle noise through the attention mechanism, and output the target detection results; Step S2 specifically includes the following steps: S21. The preprocessed SAR image first undergoes preliminary feature mapping and dimension adjustment by the initial Conv module to generate an initial feature map; S22. Initial feature map input: Perform lightweight feature extraction on the first LSConv. S221, Large Kernel Perception Submodule Processing: For the initial feature map H×W×C, firstly, the number of channels is reduced to C / 2 by 1×1 pointwise convolution to reduce the amount of computation; then, 7×7 large kernel depthwise separable convolution is applied to extract contextual information in the perception region to expand the receptive field; finally, 1×1 convolution is used to generate context-adaptive weights, and at the same time, BN layer normalization and ReLU activation are used to enhance the non-linear expression of features. S222, Small Kernel Aggregation Submodule Processing: The feature map channels input to the small kernel aggregation submodule are divided into G groups, with C / G channels in each group sharing aggregation weights; the context-adaptive weights generated in step S221 are used to dynamically aggregate the features in the small kernel neighborhood; then, feature aggregation is performed through 3×3 small kernel dynamic convolution. The grouping mechanism further reduces the computational cost, enhances the model's responsiveness to key regions, and outputs the aggregated feature map; S23. Input the aggregated feature map output by the first LSConv into the first Rep_C3k2 for feature enhancement and fusion; S231, the channel pre-aggregation submodule aggregates the channel information of the aggregated feature map and compresses the number of channels through 1×1 convolution, reducing the amount of subsequent computation; S232. The feature segmentation module divides the feature map output by the channel pre-aggregation module into two parts, A and B, according to the number of channels. Part A retains the shallow basic features and passes them to the feature splicing and fusion module; Part B is input to the multi-branch processing module. S233. The multi-branch processing submodule uses the RepConv multi-branch structure to process part B: during the training phase, deep features are extracted through parallel paths; during the inference phase, the multi-branch structure is equivalent to a single-branch structure through structural reparameterization, and the deep processing feature map is output. S234. The feature splicing and fusion submodule performs a second segmentation on the deep-processed feature map to obtain part C and part D. Then, parts A, C, and D are spliced ​​together according to the channel dimension to output the fused feature map. S24. Repeat steps S22 and S23 to generate fused feature maps at different scales, and finally output a multi-scale feature set.

5. The SAR image target detection method according to claim 4, characterized in that: Step S3 specifically includes the following steps: S31. Divide the multi-scale feature maps in the multi-scale feature set into high-level feature maps, mid-level feature maps and low-level feature maps according to resolution and semantic information. S32. Top-down path fusion: Process high-level feature maps in multi-scale feature sets to supplement the semantic discrimination ability of mid-level feature maps. S321, Upsampling Adaptation: Perform an upsampling operation on the high-level feature map to increase its resolution to be consistent with the mid-level feature map; S322, Channel stitching: The upsampled high-level feature map and the mid-level feature map are stitched together according to the channel dimension to obtain a preliminary fused feature map; S323. Input the preliminary fused feature map into the MHMC_C3k2 submodule to perform multi-scale feature enhancement and ground object interference suppression; S3231, Feature Branch Segmentation: The preliminary fused feature map is divided into branch a and branch b according to the number of channels by feature branch segmentation sub-unit. Branch a is passed to the feature splicing and channel aggregation unit, and branch b is passed to the multi-scale convolution processing sub-unit. S3232, Multi-scale Convolution Extraction: The sub-unit is processed by multi-scale convolution. First, the dimensions of the b branch are rearranged. Then, the rearranged b branch is divided into 3 sub-branches according to the number of channels. The 3 sub-branches are convolved with 3×3 convolution kernel, 5×5 convolution kernel, and 7×7 convolution kernel respectively to generate 3 sets of feature subsets containing information at different scales. S3233, Dimensional Rearrangement and Merging: The dimensional rearrangement and merging unit rearranges the three feature subsets and then merges them according to the channel dimension to obtain a merged feature map that integrates multi-scale information. S3234, Feature Concatenation and Aggregation: Through the feature concatenation and channel aggregation unit, the a branch and the merged feature map are concatenated by channel, and then a 1×1 pointwise convolutional layer is used to perform channel aggregation, finally outputting the optimized mid-layer fused feature map; S33, Bottom-up path fusion: Process the optimized mid-layer fusion feature map output by S32 to supplement the semantic information of the bottom-layer feature map in the multi-scale feature set; S331, Dimension Adjustment: A 3×3 convolutional layer is used to adjust the channel dimension and resolution of the optimized middle-layer fusion feature map so that its number of channels and resolution match the bottom-layer feature map; S332, Channel stitching: The optimized middle-layer fusion feature map after dimension adjustment is stitched together with the bottom-layer feature map according to the channel dimension to obtain the secondary fusion feature map; Secondary optimization of submodules S333 and MHMC_C3k2: Repeat step S323, and finally output the optimized small-scale fused feature map; S34. Individual optimization of high-level feature maps and three-scale integration: Individual optimization of high-level feature maps output by the backbone feature extraction network; S341, MHMC_C3k2 submodule optimization: Input the high-level feature map into the MHMC_C3k2 submodule, repeat S323, and output the optimized large-scale fused feature map; S342, Integration of Three-Scale Optimized Feature Maps: The optimized mid-scale fusion feature map output by S32, the optimized small-scale fusion feature map output by S33, and the optimized large-scale fusion feature map output by S341 are integrated to form a three-scale optimized feature map.

6. An electronic device, characterized in that, It includes a processor and a memory, wherein the processor is configured to perform the SAR image target detection method of claim 4 or 5.

7. A storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, the SAR image target detection method of claim 4 or 5 is implemented.

Images