CN115565066A - A Transformer-based ship target detection method in SAR images - Google Patents

Abstract

The invention discloses a Transformer-based SAR image ship target detection method. For small-scale SAR ship targets, the Transformer is used as a backbone network to fuse effective ship information through a deformable attention mechanism to improve detection accuracy. Firstly, Patch division is performed on the input original image. The image after patch division is input into the four-stage feature extraction backbone network composed of Transformer, and four different scale features from shallow to deep are obtained. The features of four different scales are input into the feature pyramid network for feature fusion, and the fusion features of five different scales from shallow to deep are obtained. According to the ship position annotation of the original image, the rough edge image of the target is extracted. The shallowest fused features and coarsely extracted edge images are input into the edge-guided shape enhancement module to obtain the enhanced shallowest fused features. The enhanced shallowest fused features and the fused features of the other four scales are input into the anchor-free target detection head to obtain the target detection results.

Description

A Transformer-based ship target detection method in SAR images

technical field

The invention relates to the technical field of target detection, in particular to a Transformer-based SAR image ship target detection method.

Background technique

Synthetic Aperture Radar (SAR) is an active remote sensing system whose operating frequency is in the microwave band. SAR sensors operating on aircraft or satellites can image and monitor the ground or sea level. Compared with optical sensors, the main advantage of SAR sensors is that they can work all day and all day long, and the emitted signals have strong penetration capabilities. Penetrates clouds and fog.

SAR image data increases with the wide application of SAR sensors, and the automatic detection technology for targets in SAR images has become an important research. Existing target detection methods for SAR images can be divided into traditional methods and deep learning methods. Traditional SAR target detection methods mainly include methods based on contrast information, geometric and texture features, and statistical analysis. Benefiting from the development of deep learning and GPU computing power, deep convolutional neural network (CNN) has made a huge breakthrough in object detection. At present, the target detection technology based on convolutional neural network has become the mainstream direction in the field of target detection. The algorithm is mainly divided into two types. The first type is a two-stage target detection algorithm represented by Faster R-CNN. Generate candidate areas and then perform target detection, which has high detection accuracy but low efficiency; the second type is single-stage target detection algorithms, mainly represented by SSD (Single Shot MultiBox Detector, SSD) and YOLO (You Only Look Once, YOLO) series , this type of algorithm does not generate candidate regions, and directly performs target detection through regression. The detection efficiency is high but the accuracy is not as good as the first type of method. As a new neural network structure, Transformer provides a new way of thinking for vision tasks. Originally, Transformer was used in the field of natural language processing (NLP). It uses an acyclic network structure of encoder-decoder and self-attention mechanism to achieve the best performance in machine translation. The successful application of Transformer in the field of NLP has led relevant scholars to discuss and try its application in the field of computer vision. Some backbone networks that use Transformer instead of convolution, such as ViT and Swin Transformer, have been proven to have better performance than CNN, because Transformer's global interaction mechanism can quickly expand the effective receptive field of features.

However, there are two problems in using Transformer as the backbone network in SAR ship detection. One is that the background of maritime SAR ship images is very simple, so the global relationship modeling mechanism in Transformer will associate some redundant background information. The second is that the outline between the near-shore SAR ship target and the coast is blurred, and it is difficult to distinguish the near-shore ship target from the background. Therefore, the features extracted by Transformer need to be reconstructed with more object details to focus on SAR ship targets in similar backgrounds.

At present, there is no relevant technical solution that can solve the problem of blurred ship target outlines caused by backgrounds such as coasts, islands, and waves in SAR images, thereby reducing the detection accuracy.

Contents of the invention

In view of this, the present invention provides a Transformer-based SAR image ship target detection method, which can focus on small-scale SAR ship targets, using the local sparse information aggregation Transformer based on the Swin Transformer architecture as the backbone network, through the deformable attention The sparse attention mechanism effectively fuses the effective information of small ships to improve the detection accuracy.

In order to achieve the above object, the technical solution of the present invention comprises the following steps:

Step 1: Patch the input original image to obtain the patched image.

Step 2: Input the patched image into the four-stage feature extraction backbone network composed of local sparse information aggregation Transformer, and obtain four different scale features from shallow to deep.

Step 3: Input features of four different scales into the feature pyramid network for feature fusion between different scales, and obtain fusion features of five different scales from shallow to deep after fusion.

Step 4: According to the ship position annotation of the original image, extract the edge of the target, and obtain the rough edge image of the target.

Step 5: Input the shallowest fusion feature among the fusion features of five different scales and the roughly extracted edge image into the edge-guided shape enhancement module to obtain the enhanced shallowest fusion feature.

Step 6: Input the enhanced shallowest fusion feature and fusion features of other four scales into the target detection head without anchor frame to obtain the target detection result.

Further, the patch division of the input original image includes the following steps: divide the input image with a size of H×W×3 into 4×4 non-overlapping patches, and then the feature dimension of each patch is 4×4×3 =48, the number of patches is H/4×W/4.

Further, the four-stage feature extraction backbone network composed of local sparse information aggregation Transformer includes: the basic composition structure is based on the backbone network structure of Swin Transformer, which is divided into four stages, which are respectively recorded as stage 1, stage 2, stage 3 and In the fourth stage, the features of four different scales from shallow to deep are output in sequence.

Phase 1 includes sequentially connected linear embedding modules and dual Transformer modules; the linear embedding module is used to perform dimension transformation on the patch-divided image.

Phase two includes sequentially connected patch fusion modules and dual Transformer modules.

Phase three includes sequentially connected patch fusion modules and three sequentially connected dual Transformer modules.

Phase four includes sequentially connected patch fusion modules and dual Transformer modules.

The double Transformer module includes two Transformer modules before and after, and the front Transformer module performs the following steps: calculate the local sparse information aggregation attention map for the input features of the front Transformer module, and make residual difference between the local sparse information aggregation attention map and the input features of the front Transformer module Connect, and then residually link this result with the result after linear transformation and multi-layer perceptron to obtain the output features of the local sparse information aggregation Transformer.

The output features of the pre-Transformer module are used as the input features of the post-Transformer module.

The post-Transformer module performs the following steps: After sampling and transforming the input features of the post-Transformer module, calculate the local sparse information aggregation attention map for the features after sampling transformation, and perform residual difference between the local sparse information aggregation attention map and the input features of the post-Transformer module Connect, and then residually link this result with the result after linear transformation and multi-layer perceptron to obtain the output features of the local sparse information aggregation Transformer.

The sampling transformation includes: passing the input features of the post-Transformer module through the convolution layer to obtain data-based sampling values; using the data-based sampling values to perform bilinear interpolation sampling on the input features of the post-Transformer module to obtain the features after sampling transformation.

Among them, the local sparse information aggregation attention map is calculated, and the input features of the former Transformer module or the input features of the post-Transformer module after sampling transformation are used as the current input feature A; the following steps are adopted:

S1: Divide the current input feature A into windows of the same size and do not overlap each other to obtain a window-divided feature map.

S2: Input the window division feature map into the offset generation network to obtain the offset matrix required for calculating deformable attention.

S3: Linearly transform the window-divided feature map to obtain the value matrix required for calculating deformable attention.

S4: Linearly transform the window-divided feature map to obtain the attention weight matrix required for calculating deformable attention.

S5: Use the offset matrix to perform bilinear interpolation sampling on the value matrix, and use the attention weight matrix to weight and sum the sampling results, and then perform linear transformation to obtain the local sparse information aggregation attention map.

Further, the local sparse information aggregates the attention map, which is expressed as DeformAttn(z _q ,p _q ,x):

Among them, z _q and x are two representations of input features, p _q is any reference point on the feature, Δp _mqk is the offset, A _mpk is the attention weight, W _m is the learnable weight, W′ _m is The transposition of W _m , M is the number of attention heads, and K is the number of sampling points.

Further, the patch fusion module includes: taking the values of the same position of each calculation area in the output features of the dual Transformer modules to form a new patch and connecting them to obtain the features after 2 times downsampling.

Further, the input of the feature pyramid network is features C1~C4 of four different scales from shallow to deep;

The feature pyramid network performs the following processing on features C1~C4 to obtain fusion features P1~P5 of five different scales from shallow to deep after fusion: C4 directly obtains P4, P4 is down-sampled to obtain P5, P4 is up-sampled and fused with C3 to obtain P3, P3 is up-sampled and fused with C2 to get P2, and P2 is up-sampled and fused with C1 to get P1.

Further, the edge-guided shape enhancement module includes:

The shallowest fusion feature output by Transformer is fused with the rough extraction edge image, and input to the sobel edge extraction operator to obtain the target edge prediction map.

Computes the weighted cross-entropy loss between the predicted map of the target edge and the ground truth map of the target edge.

The normalized target edge prediction map is used to weight the shallowest fusion feature output by the Transformer and fuse with it, and input the convolutional network to obtain the target shape prediction map.

Calculate the binary classification loss and Dice loss between the target shape prediction map and the target binary segmentation real map. Based on the above losses, optimize the convolutional network parameters and enhance the features.

Further, the weighted cross-entropy loss between the target edge prediction map and the target edge real map, which is expressed as Loss _ce :

Among them, y _i is the edge pixel value of the target edge-truth map, p _i is the probability that the i-th pixel is classified as an edge pixel, conversely, 1-y _i is the background pixel value of the target edge-truth map, 1-p _i is the probability that the i-th pixel is classified as a background pixel; N _c is the number of edge pixels in the target edge-truth image; N is all pixels in the target edge-truth image.

Further, the binary classification loss and Dice loss between the target shape prediction map and the target binary segmentation real map are expressed as:

Among them, Loss _se is the binary classification loss between the target shape prediction map and the target binary segmentation real map; Loss _Dice is the Dice loss between the target shape prediction map and the target binary segmentation real map; y _i is the target edge real map The edge pixel value of , p _i is the probability that the i-th pixel is classified as an edge pixel, conversely, 1-y _i is the background pixel value of the real image of the target edge, and 1-p _i is the i-th pixel is classified is the probability of the background pixel; α determines the weight of the two losses.

Further, the object detection head without anchor box includes the detection head of FCOS, and the results of object detection classification and regression are obtained.

Beneficial effect:

1. The SAR image ship target detection method based on the explicit edge-guided local sparse information aggregation Transformer provided by the present invention, for small-scale SAR ship targets, a local sparse information aggregation Transformer based on the Swin Transformer architecture is proposed as the backbone network , to effectively fuse the effective information of small ships through a sparse attention mechanism of deformable attention.

2. The SAR image ship target detection method based on explicit edge-guided local sparse information aggregation Transformer provided by the present invention, when replacing the self-attention mechanism with a deformable attention mechanism, combines a data-dependent offset generator, In order to obtain more prominent small SAR ship target characteristics.

3. The SAR image ship target detection method based on the explicit edge-guided local sparse information aggregation Transformer provided by the present invention proposes an explicit edge-guided shape enhancement module, which can more effectively enhance the fuzzy outline of the features extracted by the Transformer. SAR ships and distinguish them from background interference.

Description of drawings

1 is a schematic flow diagram of a method for detecting a ship target in a SAR image based on an explicit edge-guided local sparse information aggregation Transformer according to Embodiment 1 of the present invention;

FIG. 2 is a schematic diagram of a feature extraction network structure according to Embodiment 1 of the present invention;

FIG. 3 is a schematic structural diagram of a local sparse information aggregation Transformer module according to Embodiment 1 of the present invention;

FIG. 4 is a schematic structural diagram of an edge-guided shape enhancement module according to Embodiment 1 of the present invention;

FIG. 5 is a schematic structural diagram of a method for detecting a ship target in a SAR image based on an explicit edge-guided local sparse information aggregation Transformer according to Embodiment 1 of the present invention.

detailed description

The present invention will be described in detail below with reference to the accompanying drawings and examples.

This embodiment provides a method for detecting a ship target in a SAR image based on an explicit edge-guided local sparse information aggregation Transformer, the process of which is shown in FIG. 1 . First, the non-overlapping 4*4 pixel area in the input image is divided into a patch, and the image divided by the patch is obtained, and it is input into a four-stage feature extraction backbone network composed of a local sparse information aggregation Transformer, which includes a shallow The features of four different scales from deep to deep; after that, the features are input into the feature pyramid network for feature fusion between different scales, and the features of five different scales from shallow to deep after fusion are obtained; at the same time, according to the ship of the original image Position labeling to get the rough extracted edge image of the target; then, input the shallowest feature and rough extracted edge image into the edge-guided shape enhancement module to obtain the enhanced shallowest feature; finally, input the feature into the anchor box-free The target detects the head and obtains the target detection result.

The specific implementation process of this program includes the following steps:

Step 1: Perform patch division on the input original image to obtain the patched image.

The non-overlapping 4×4 pixel area in the input image is divided into a patch, that is, the input image with a size of H×W×3 is divided into 4×4 non-overlapping patches, and the feature dimension of each patch is 4 ×4×3=48, the quantity of patch is H/4×W/4, obtains the subsequent structure of image input after patch division;

Step 2: Input the patched image into the four-stage feature extraction backbone network composed of local sparse information aggregation Transformer, and obtain four different scale features from shallow to deep;

The improved Swin Transformer feature extraction network is shown in Figure 2.

The basic composition structure is based on the backbone network structure of Swin Transformer, which is divided into four stages, which are recorded as stage 1, stage 2, stage 3 and stage 4 in sequence, and the features of four different scales from shallow to deep are output in sequence;

Phase 1 includes a sequentially connected linear embedding module and a double Transformer module; the linear embedding module is used to perform dimension transformation on the patch-divided image;

Phase 2 includes sequentially connected patch fusion modules and dual Transformer modules;

Phase three includes sequentially connected patch fusion modules and three sequentially connected dual Transformer modules;

Phase four includes sequentially connected patch fusion modules and dual Transformer modules;

The dual Transformer module includes two Transformer modules before and after, and the former Transformer module performs the following steps:

Calculate the local sparse information aggregation attention map for the input features of the pre-Transformer module, perform residual connection between the local sparse information aggregation attention map and the input features of the pre-Transformer module, and then combine this result with the result after linear transformation and multi-layer perceptron Perform residual linking to obtain the output features of the local sparse information aggregation Transformer;

The output features of the former Transformer module are used as the input features of the post-Transformer module;

The post-Transformer module performs the following steps:

After sampling and transforming the input features of the post-Transformer module, the local sparse information aggregation attention map is calculated for the features after the sampling transformation, and the local sparse information aggregation attention map is residually connected with the input features of the post-Transformer module, and then the result is combined with After linear transformation and multi-layer perceptron, the results are residually linked to obtain the output features of the local sparse information aggregation Transformer.

The sampling transformation includes: passing the input features of the post-Transformer module through the convolution layer to obtain data-based sampling values; using the data-based sampling values to perform bilinear interpolation sampling on the input features of the post-Transformer module to obtain the features after sampling transformation.

Among them, the local sparse information aggregation attention map is calculated, and the input features of the former Transformer module or the input features of the post-Transformer module after sampling transformation are used as the current input feature A; the following steps are adopted:

S1: Divide the current input feature A into windows of the same size and do not overlap each other to obtain a window-divided feature map.

S2: Input the window division feature map into the offset generation network to obtain the offset matrix required for calculating deformable attention.

S3: Linearly transform the window-divided feature map to obtain the value matrix required for calculating deformable attention.

S4: Linearly transform the window-divided feature map to obtain the attention weight matrix required for calculating deformable attention.

S5: Use the offset matrix to perform bilinear interpolation sampling on the value matrix, and use the attention weight matrix to weight and sum the sampling results, and then perform linear transformation to obtain the local sparse information aggregation attention map.

The structure of the Local Sparse Information Aggregation Transformer module is shown in Figure 3.

The calculation formula of deformable attention is:

Among them, DeformAttn(z _q ,p _q ,x) is the attention map, z _q and x are two representations of input features, p _q is any reference point on the feature, Δp _mqk is the offset, A _mpk is the attention Weight, W _m is the learnable weight, W′ _m is the transpose of W _m , M is the number of attention heads, and K is the number of sampling points.

The above attention map is residually connected to the input features, and then the result is residually linked with the result after linear transformation and multi-layer perceptron to obtain the output features of the local sparse information aggregation Transformer.

Due to the lack of information interaction between windows in every two consecutive Transformer modules, it is necessary to sample and transform the feature map obtained by the previous Transformer module, that is, to input the feature into the convolutional layer to obtain a data-based sampling value, and then use the sampling value to this The feature is sampled by bilinear interpolation, and then the calculation of the latter Transformer module is performed.

The linear embedding module performs dimension transformation on the patched image.

The patch fusion module takes the value of the same position of each calculation area in the output feature of Transformer to form a new patch and connects it to obtain the feature after 2 times downsampling.

The output of the feature extraction network is features of four different scales from shallow to deep.

Step 3: Input features of four different scales into the feature pyramid network for feature fusion between different scales, and obtain fusion features of five different scales from shallow to deep after fusion.

As shown in the feature fusion pyramid structure in Figure 5, the input of the feature pyramid network is four different scale features C1~C4 from shallow to deep; the feature pyramid network performs the following processing on the features C1~C4 to obtain the fusion from shallow to deep Fusion features of five different scales P1～P5: C4 directly obtains P4, P4 is down-sampled to obtain P5, P4 is up-sampled and fused with C3 to obtain P3, P3 is up-sampled and fused with C2 to obtain P2, P2 is up-sampled and fused with C1 to obtain P1.

Step 4: According to the ship position annotation of the original image, extract the edge of the target, and obtain the rough edge image of the target.

Step 5: Input the shallowest fusion feature among the fusion features of five different scales and the roughly extracted edge image into the edge-guided shape enhancement module to obtain the enhanced shallowest fusion feature.

Input the shallowest feature of the five different scale features and the rough extraction edge image into the edge-guided shape enhancement module, that is, first fuse the shallowest feature with the rough extraction edge image, and input the target using the sobel edge extraction operator edge prediction map, and then calculate the weighted cross-entropy loss between the target edge prediction map and the target edge real map; then use the normalized target edge prediction map to weight the shallowest feature and fuse it, and input the convolution The network obtains the target shape prediction map, and calculates the binary classification loss and Dice loss between the target shape prediction map and the target binary segmentation real map; finally, based on the above losses, the parameters are optimized by training the network to enhance the features.

The weighted cross-entropy loss between the target edge prediction map and the target edge real map is calculated as:

Among them, y _i is the edge pixel value of the target edge-truth map, p _i is the probability that the i-th pixel is classified as an edge pixel, conversely, 1-y _i is the background pixel value of the target edge-truth map, 1-p _i is the probability that the i-th pixel is classified as a background pixel; N _c is the number of edge pixels in the target edge-truth image; N is all pixels in the target edge-truth image.

The calculation formulas of binary classification loss and Dice loss between the target shape prediction map and the target binary segmentation real map are:

Among them, Loss _se is the binary classification loss between the target shape prediction map and the target binary segmentation real map; Loss _Dice is the Dice loss between the target shape prediction map and the target binary segmentation real map; y _i is the target edge real map The edge pixel value of , p _i is the probability that the i-th pixel is classified as an edge pixel, conversely, 1-y _i is the background pixel value of the real image of the target edge, and 1-p _i is the i-th pixel is classified is the probability of the background pixel; α determines the weight of the two losses.

Step 6: Input the enhanced shallowest fusion feature and fusion features of other four scales into the target detection head without anchor frame to obtain the target detection result.

As shown in the detection head in Figure 5, the enhanced shallowest layer and the features of the other four scales are input into the target detection head of the anchor-free frame of FCOS, and the target detection classification and regression results are obtained.

Those skilled in the art should understand that each step and each module of the present invention described in the above embodiments can be realized by general-purpose computing hardware and software, and they can be concentrated on a single computing device, or distributed among multiple computing devices. formed on the network. Modules implemented with executable program code may choose to store them in random access memory (RAM), internal memory, read-only memory (ROM), hard disk, removable disk, CD-ROM, etc. in any other form known in the technical field in computer storage media. In some cases, the steps shown or described may be performed in a different order than here, or they may be fabricated separately as individual integrated circuit modules, or multiple of them or steps may be fabricated as a single integrated circuit module to realise. Therefore, the present invention is not limited to any specific combination of hardware and software.

To sum up, the above are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. a Transformer-based SAR image ship target detection method, is characterized in that, comprises the steps: Step 1: Patch the input original image to obtain the patched image; Step 2: Input the image divided by the patch into the four-stage feature extraction backbone network composed of local sparse information aggregation Transformer, and obtain four different scale features from shallow to deep; Step 3: Input the features of the four different scales into the feature pyramid network for feature fusion between different scales, and obtain fusion features of five different scales from shallow to deep after fusion; Step 4: According to the ship position annotation of the original image, extract the edge of the target, and obtain the roughly extracted edge image of the target; Step 5: input the shallowest fusion feature among the five fusion features of different scales and the rough extracted edge image into the edge-guided shape enhancement module to obtain the enhanced shallowest fusion feature; Step 6: Input the enhanced fusion features of the shallowest layer and the fusion features of other four scales into the target detection head without anchor frame to obtain the target detection result. 2. the SAR image ship target detection method based on Transformer as claimed in claim 1, is characterized in that, the described original image of input is carried out Patch division and comprises the steps: Divide the input image of size H×W×3 into 4×4 non-overlapping patches, then the feature dimension of each patch is 4×4×3=48, and the number of patches is H/4×W/4. 3. the Transformer-based SAR image ship target detection method as claimed in claim 1, is characterized in that, the four-stage feature extraction backbone network composed of local sparse information aggregation Transformer comprises: The basic composition structure is based on the backbone network structure of Swin Transformer, which is divided into four stages, which are recorded as stage 1, stage 2, stage 3 and stage 4 in sequence, and the features of four different scales from shallow to deep are output in sequence; Phase one comprises a linear embedding module and a double Transformer module connected in sequence; the linear embedding module is used to perform dimension transformation on the image divided by the patch; Phase 2 includes sequentially connected patch fusion modules and dual Transformer modules; Phase three includes sequentially connected patch fusion modules and three sequentially connected dual Transformer modules; Phase four includes sequentially connected patch fusion modules and dual Transformer modules; Described double Transformer module comprises two Transformer modules before and after, wherein former Transformer module performs the following steps: Calculate the local sparse information aggregation attention map for the input features of the pre-Transformer module, perform residual connection between the local sparse information aggregation attention map and the input features of the front Transformer module, and then compare the result with the linear transformation and multi-layer perceptron. The results of the residual link are obtained to obtain the output features of the local sparse information aggregation Transformer; The output features of the former Transformer module are used as the input features of the post-Transformer module; The post-Transformer module performs the following steps: After sampling and transforming the input features of the post-Transformer module, calculate the local sparse information aggregation attention map for the features after the sampling transformation, and perform residual connection between the local sparse information aggregation attention map and the input features of the post-Transformer module, and then connect the The result is residually linked with the result after linear transformation and multi-layer perceptron, and the output features of the local sparse information aggregation Transformer are obtained; The sampling transformation includes: passing the input feature of the post-Transformer module through a convolution layer to obtain a data-based sampling value; using the data-based sampling value to perform bilinear interpolation sampling on the input feature of the post-Transformer module to obtain a post-sampling transformation. Characteristics; Among them, the local sparse information aggregation attention map is calculated, and the input features of the former Transformer module or the input features of the post-Transformer module after sampling transformation are used as the current input feature A; the following steps are adopted: S1: Divide the current input feature A into windows of the same size and do not overlap each other to obtain a feature map of the window division; S2: Input the window division feature map into the offset generation network to obtain the offset matrix required for calculating the deformable attention; S3: performing linear transformation on the window-divided feature map to obtain a value matrix required for calculating deformable attention; S4: performing linear transformation on the window division feature map to obtain an attention weight matrix required for calculating deformable attention; S5: Use the offset matrix to perform bilinear interpolation sampling on the value matrix, and use the attention weight matrix to weight and sum the sampling results, and then perform linear transformation to obtain a local sparse information aggregation attention map. 4. The Transformer-based SAR image ship target detection method according to any one of claims 1 to 3, wherein the local sparse information aggregation attention map is expressed as DeformAttn(z _q , p _q , x) :

Among them, z _q and x are two representations of input features, p _q is any reference point on the feature, Δp _mqk is the offset, A _mpk is the attention weight, W _m is the learnable weight, W′ _m is the transpose of W _m , M is the number of attention heads, and K is the number of sampling points. 5. the SAR image ship target detection method based on Transformer as claimed in claim 3, is characterized in that, described patch fusion module comprises: get the value spelling of the identical position of each computation area in the output feature of described double Transformer module Form a new patch and connect it to get the features after 2 times downsampling. 6. the SAR image ship target detection method based on Transformer as claimed in claim 1, is characterized in that, the input of described feature pyramid network is the feature C1～C4 of four different scales from shallow to deep; The feature pyramid network performs the following processing on features C1~C4 to obtain fusion features P1~P5 of five different scales from shallow to deep after fusion: C4 directly obtains P4, P4 is down-sampled to obtain P5, P4 is up-sampled and fused with C3 to obtain P3, P3 is up-sampled and fused with C2 to get P2, and P2 is up-sampled and fused with C1 to get P1. 7. the SAR image ship target detection method based on Transformer as claimed in claim 1, is characterized in that, the shape enhancement module of described edge guidance comprises: Said Transformer outputs the shallowest fusion feature and said rough extraction edge image fusion, and imports the sobel edge extraction operator to obtain the target edge prediction map; Calculate the weighted cross-entropy loss between the target edge prediction map and the target edge real map; Using the normalized target edge prediction map to weight the Transformer output shallowest fusion feature and then fuse with it, and input the convolution network to obtain the target shape prediction map; Calculate the binary classification loss and Dice loss between the target shape prediction map and the target binary segmentation real map. Based on the above losses, optimize the convolutional network parameters and enhance the features. 8. the SAR image ship target detection method based on Transformer as claimed in claim 11, is characterized in that, the weighted cross-entropy loss between described target edge prediction map and target edge real map, it is expressed as Loss _ce :

Among them, y _i is the edge pixel value of the target edge-truth map, p _i is the probability that the i-th pixel is classified as an edge pixel, conversely, 1-y _i is the background pixel value of the target edge-truth map, 1-p _i is the probability that the i-th pixel is classified as a background pixel; N _c is the number of edge pixels in the target edge-truth image; N is all pixels in the target edge-truth image. 9. the SAR image ship target detection method based on Transformer as claimed in claim 11, is characterized in that, the binary classification loss and the Dice loss between the target shape prediction map and the target binary segmentation real map, it is expressed as :

Among them, Loss _se is the binary classification loss between the target shape prediction map and the target binary segmentation real map; Loss _Dice is the Dice loss between the target shape prediction map and the target binary segmentation real map; y _i is the target edge real map The edge pixel value of , p _i is the probability that the i-th pixel is classified as an edge pixel, conversely, 1-y _i is the background pixel value of the real image of the target edge, and 1-p _i is the i-th pixel is classified is the probability of the background pixel; α determines the weight of the two losses. 10. The Transformer-based SAR image ship target detection method according to claim 1, wherein the target detection head without anchor frames comprises the detection head of FCOS, and obtains the results of target detection classification and regression.

Images