Lightweight SAR image ship detection system and method based on strip pruning

Abstract

The present invention relates to a lightweight SAR image ship detection method based on strip pruning, and belongs to the field of image processing technology. A lightweight residual convolutional network based on strip convolution is used to extract feature maps of different depths; a jump connection feature pyramid network is used to fuse features of different depths to obtain fused depth feature maps of different scales; a regression model based on key points is used to predict the possibility probability of the corresponding position of each feature point and the vertical distance from the feature point position to the target edge, and finally the target position and probability are obtained. This method uses a lightweight residual convolutional network based on strip convolution to obtain a lightweight residual convolutional network through one end-to-end training, thereby improving the training efficiency and robustness of the model; strip pruning increases the pruning dimension, and further compresses the number of parameters and the amount of calculation of the model.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and relates to a ship target detection method of SAR images, in particular to a lightweight SAR image ship detection method for limited storage and computing equipment. Background Art

[0002] SAR images have the characteristics of all-day and all-weather imaging, and ship detection using SAR images is of great significance in fields such as ocean monitoring and military. However, due to the limitations of the hardware conditions of the satellite terminal environment, the SAR image ship detection method needs to reduce the parameter scale and calculation amount as much as possible while ensuring the detection performance.

[0003] "On-Board Real-Time Ship Detection in HISEA-1SAR Images Based on CFARand Lightweight Deep Learning" (Remote Sensing. IEEE, 2021: 1995-2013.) uses a dual-parameter constant false alarm rate to extract images that may contain ships, and then uses the real-time detection algorithm YOLO V4 to detect image blocks. "TWC-Net: A SAR Ship Detection Using Two-Way Convolution and Multiscale Feature Mapping" (Remote Sensing. IEEE, 2021: 2558-2574.) designs a dual-channel convolutional network that can extract feature maps of different scales and semantic information to process targets of different scales. These two methods use lightweight single-step detection algorithms, which can basically achieve real-time detection, but because the convolutional network extracts deep semantic information, its network parameters are huge, which is not conducive to storage and calculation. "SAR Ship Target Detection Method Based on Network Pruning and Knowledge Distillation" (CN112308019) designed a DC-ACMYOLOV3 detection algorithm, used regularization loss to make parameters sparse, then pruned the network channel, and finally used knowledge distillation to tune the pruned network. This method has independent pruning and training processes, and cannot achieve end-to-end training. It only prunes at the channel layer and does not completely solve the problem of parameter redundancy. Summary of the invention

[0004] Technical issues to be solved

[0005] In order to further reduce the algorithm parameter scale and calculation amount and ensure the robustness and detection performance of the detection algorithm, the present invention provides a lightweight SAR image ship detection system and method based on strip pruning.

[0006] Technical Solution

[0007] A lightweight SAR image ship detection system based on stripe pruning is characterized by comprising a lightweight residual convolutional network based on stripe convolution, a skip connection feature pyramid network and a regression model based on key points; feature maps of different depths are extracted by using a lightweight residual convolutional network based on stripe convolution; feature maps of different depths are extracted by using a lightweight residual convolutional network based on stripe convolution, and the feature maps are passed into a skip connection feature pyramid network to obtain a fused feature map that fully integrates deep features and shallow features, and a regression model based on key points is used to directly predict the score and prediction box position of each feature point.

[0008] A further technical solution of the present invention: The lightweight residual convolution network based on strip convolution is improved from ResNet50, specifically, the standard 3*3 convolution kernel in the ResNet50 network is replaced by a strip convolution kernel and a convolution kernel skeleton matrix, the strip convolution layer is composed of a standard convolution layer and a convolution kernel skeleton matrix, and during calculation, the convolution kernel is dot-multiplied with the convolution kernel skeleton matrix to form a general convolution kernel for convolution operation; the convolution kernel skeleton matrix is ​​composed of learnable parameters, each parameter corresponds to a strip vector on the convolution kernel, the larger the parameter, the greater the weight of the strip when extracting features, and during training, when the parameter is lower than the threshold, the strip is removed to achieve strip pruning.

[0009] A further technical solution of the present invention is that the skip connection feature pyramid network fuses the extracted 4-layer feature maps twice from bottom to top and from top to bottom layer by layer to obtain 3-layer fused feature maps of different sizes.

[0010] A further technical solution of the present invention is as follows: the key point-based regression model is divided into a key point confidence regression subnetwork and a key point distance regression subnetwork. The key point confidence regression subnetwork predicts a score map of the key point, which is sequentially composed of three 3x3 convolutional layers, a 1x1 convolutional layer and a sigmoid function. The score map output by the network has the same size as the input feature map, and the value of the element on the score map indicates the probability that the key point of the ship target exists at the corresponding position point of the original image; the position regression subnetwork predicts the distance from the key point to the four sides of the prediction box, and the number of channels of the output feature map is 4, representing the distance prediction amount (t1, t2, t3, t4) of the key point. The prediction box position is located by the key point position and the distance from the key point to the four sides of the prediction box. The distance between the key point and the prediction box edge is calculated by the distance prediction amount and the learnable scale parameter, and the formula is as follows:

[0011] d i =exp(l j )×exp(t i ), for i=1,2,3,4

[0012] Among them, d i is the distance prediction quantity, l j is the size parameter of the science department, t i is the network distance prediction quantity.

[0013] A lightweight SAR image ship detection method based on strip pruning, characterized by the following steps:

[0014] Step 1: Extract multi-scale feature maps using a lightweight residual convolutional network based on strip convolution

[0015] The SAR image block is passed into the lightweight residual convolution network based on strip convolution for feature extraction. After 50 strip convolution layers, 4 feature maps with different depths and sizes are extracted. The sizes of the feature maps are [256, w / 4, h / 4], [512, w / 8, h / 8], [1024, w / 16, h / 16], [2048, w / 32, h / 32] from large to small, where w and h are the original image sizes.

[0016] Step 2: Skip connection feature pyramid network to fuse different depth features

[0017] The four feature maps of different sizes and depths obtained in step 1 are passed into the skip connection feature pyramid network to fuse the features of different layers. The skip connection feature pyramid network performs two feature fusions from bottom to top and top to bottom. The bottom-up fusion is similar to the feature pyramid network. The network first converts the number of channels of feature maps of different layers from 256, 512, 1024, and 2048 to 256 through a 1*1 convolution layer, and then the feature map is upsampled by 2 times from top to bottom through the nearest neighbor interpolation method and added to the feature map of the next layer element by element. The added feature map is then passed through a 3*3 convolution layer to obtain a fused feature map; the fused feature map is then downsampled layer by layer from bottom to top to the original The feature map of the same size obtained by 1*1 convolution jump link and the fused feature map of the previous layer are added element by element, and then the feature map of the three added images is passed through a 3*3 convolution layer to obtain the secondary fused feature map [256,w / 4,h / 4], [256,w / 8,h / 8], [256,w / 16,h / 16], where w and h are the original image sizes; the feature map is assigned to predict target ships of different sizes according to the size, and the target with the maximum size of [0,64] is predicted by the large-scale feature map; the target with the maximum size of [64,128] is predicted by the medium-scale feature map; the target with the maximum size of [128,INF] is predicted by the small-scale feature map;

[0018] Step 3: Prediction based on keypoint regression model

[0019] The 3-layer fusion feature maps of different scales are trained based on the key point regression model to produce the confidence prediction score map and position prediction matrix. The points with scores greater than 0.5 in the score map are selected as the prediction key points, and the four-side distance is calculated through the position prediction matrix and the calculation formula to finally get the prediction box position. Finally, the soft non-maximum suppression algorithm is used to exclude overlapping prediction boxes to get the final prediction result.

[0020] A further technical solution of the present invention is as follows: the skip connection feature pyramid network uses the Kaiming initialization method to initialize the network parameters.

[0021] A further technical solution of the present invention is as follows: the key point-based regression model uses the Kaiming initialization method to initialize network parameters.

[0022] Beneficial Effects

[0023] The present invention proposes a lightweight SAR image ship detection system and method based on strip pruning, which utilizes a lightweight residual convolutional network based on strip convolution to extract feature maps of different depths; utilizes a jump connection feature pyramid network to fuse features of different depths to obtain fused depth feature maps of different scales; utilizes a key point-based regression model to predict the possibility probability of the corresponding position of each feature point and the vertical distance from the feature point position to the target edge, and finally obtains the target position and probability.

[0024] This method uses a lightweight residual convolutional network based on striped convolution to obtain a lightweight residual convolutional network through one end-to-end training, which improves the training efficiency and robustness of the model; striped pruning increases the pruning dimension and further compresses the number of parameters and calculations of the model. After extracting features of different depths, the method uses a jump connection feature pyramid network to fully combine the shallow and deep features of the target, which is conducive to the detection of ship targets of different sizes. In the prediction stage, a regression model based on key points is used to greatly reduce the number of predicted candidate boxes and reduce the amount of calculation; learnable equal-scale parameters are used to avoid data deviations caused by preset scale parameters. Its specific advantages are as follows:

[0025] (1) A lightweight residual convolutional network based on striped convolution is constructed. During the training process, the network weights are optimized and the channels and shapes of the convolution kernels are pruned. A lightweight residual network for extracting image features can be obtained at one time, avoiding network changes caused by pruning after training and eliminating the need for post-pruning tuning training. End-to-end training and pruning improve the robustness and training efficiency of the model.

[0026] (2) Strip pruning implements pruning on the channel dimension and shape of the convolution kernel, increasing the pruning dimension. Compared with channel pruning, it can further reduce the number of model parameters and the amount of computation. Strip pruning also implements convolution kernel shape learning, allowing the convolution kernel to more effectively extract the required features and improve the efficiency of the model in extracting features.

[0027] (3) A skip connection feature pyramid network was designed. Through skip connections, features at different levels were fused twice from bottom to top and from top to bottom using a small number of convolution operations. This enabled the fused feature map to contain both deep semantic information and shallow location information, which is beneficial for SAR ship target detection of different scales under complex backgrounds.

[0028] (4) The model adopts a regression model based on key points, which reduces the amount of computation caused by the anchor box-based detection method to calculate a large number of anchor boxes, which is conducive to the deployment and operation of the algorithm under limited hardware conditions. At the same time, setting the anchor box size hyperparameter requires strong prior knowledge. Different data sets need to set different anchor box sizes according to the target. Changes in target size will also cause a significant decrease in detection performance. The key point-based predictor adopts a learnable scale parameter, which eliminates the step of setting the anchor box size parameter and improves the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0030] Figure 1 This is a flow chart of the lightweight SAR image ship detection method based on strip pruning;

[0031] Figure 2 It is the feature extraction network structure diagram. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0033] The present invention proposes a lightweight SAR image ship detection system based on strip pruning, which includes a lightweight residual convolution network based on strip convolution, a skip connection feature pyramid network and a regression model based on key points. The lightweight residual convolution network based on strip convolution is used to extract feature maps of different depths, and the feature maps are passed into the skip connection feature pyramid network to obtain a fusion feature map that fully integrates deep features and shallow features. The regression model based on key points is used to directly predict the score and prediction box position of each feature point. The details are as follows:

[0034] (1) A lightweight residual convolutional network based on striped convolution is used to extract feature maps of different depths. ResNet50 is improved by replacing the standard 3*3 convolution kernel in the network with a striped convolution kernel and a convolution kernel skeleton matrix. The model training process simultaneously learns the striped convolution kernel and the convolution kernel skeleton matrix, and when the element value of the convolution kernel skeleton matrix is ​​lower than the threshold, the corresponding striped convolution kernel is pruned, and finally a trained lightweight network is obtained. In the test phase, the pruned model is directly used for feature extraction to obtain feature maps of different depths.

[0035] (2) The skip connection feature pyramid network is used to fuse features of different depths. The four-layer feature maps extracted by the lightweight residual convolutional network based on strip convolution are fused twice from bottom to top and from top to bottom to obtain three-layer fused feature maps of different sizes.

[0036] (3) The key point-based regression model is used to predict the probability of the corresponding position of each feature point and the vertical distance from the feature point position to the target edge. The feature vector of each point on the feature map is input into the prediction network to obtain the score and distance prediction value of the point. The vertical distance from the point to the prediction box can be calculated based on the distance prediction value and the trained learnable scale parameter. Combined with the corresponding image position of the feature point, the final prediction box position is finally obtained.

[0037] (4) The model is trained on the public SAR image ship dataset SSDD and the self-built dataset TSSD. The residual network trained on ImageNet is imported into the lightweight residual convolution network based on strip convolution, and the convolution kernel skeleton matrix, jump connection feature pyramid network and key point-based regression model are initialized. The loss function consists of three parts: classification loss function, distance regression loss function and L1 regularization loss of the convolution kernel skeleton matrix. The classification loss function and distance regression loss function use the cross entropy loss function and IoU loss function respectively. The stochastic gradient descent algorithm is used for model training, and pruning is completed during the training process. Finally, a lightweight SAR image ship detection system is obtained.

[0038] The present invention also proposes a lightweight SAR image ship detection method based on strip pruning, the steps are as follows:

[0039] 1. Lightweight residual convolutional network based on strip convolution to extract multi-scale feature maps

[0040] The SAR image block is passed into the lightweight residual convolution network based on strip convolution for feature extraction to obtain feature maps of different sizes. The network structure diagram is shown in Figure 2 , the steps are as follows:

[0041] (1) The lightweight residual convolution network based on strip convolution is improved from the residual network ResNet50, which changes all standard convolution layers in ResNet50 into strip convolution layers. The strip convolution layer consists of a standard convolution layer and a convolution kernel skeleton matrix. During calculation, the convolution kernel is dot-multiplied with the convolution kernel skeleton matrix to form a general convolution kernel for convolution operation. The convolution kernel skeleton matrix consists of learnable parameters, each of which corresponds to a strip vector on the convolution kernel. The larger the parameter, the greater the weight of the strip when extracting features. During the training process, when the parameter is lower than the threshold of 0.005, the strip is removed to achieve strip pruning.

[0042] (2) The convolution kernels in the strip convolution layer are initialized with the classification model trained on the ImageNet classification dataset as the pre-trained model. The convolution kernel skeleton matrix in the strip convolution layer is initialized to a learnable matrix with all element values ​​1. The loss function adds the L1 regularization term of the convolution kernel skeleton matrix to sparse the convolution kernel skeleton matrix. The L1 regularization term coefficient is 0.0005. During the training process, when an element in the convolution kernel skeleton matrix is ​​lower than the threshold of 0.005, the element is reset to 0, and the element and the corresponding strip sum are no longer updated. When the training is completed, the element below the threshold will be pruned.

[0043] (3) After the training is completed, the convolution kernel in the strip convolution layer is decomposed into 1*1 convolutions, the convolutions below the threshold are removed, and the calculation of the original convolution layer is restored according to the convolution kernel skeleton matrix, and finally a lightweight residual network after strip pruning is obtained.

[0044] (4) After the SAR image is passed into the lightweight residual network, 50 strip convolutional layers are used to extract four feature maps of different depths and sizes. The sizes of the feature maps are ([256, w / 4, h / 4], [512, w / 8, h / 8], [1024, w / 16, h / 16], [2048, w / 32, h / 32], w and h are the original image sizes) from large to small for subsequent prediction.

[0045] 2. Skip connection feature pyramid network integrates different depth features

[0046] The feature maps of different sizes and depths are passed into the skip connection feature pyramid network to fuse the features of different layers to obtain fused feature maps of different sizes. The network structure diagram is shown in the figure. Figure 2 , the steps are as follows:

[0047] (1) The jump connection feature pyramid network performs two feature fusions from bottom to top and top to bottom. The bottom-up fusion is similar to the feature pyramid network. The network first converts the number of channels of feature maps of different layers from 256, 512, 1024, and 2048 to 256 through a 1*1 convolution layer. Then the feature map is upsampled by 2 times from top to bottom through the nearest neighbor interpolation method and element-wise added to the feature map of the next layer. The added feature map is then passed through a 3*3 convolution layer to obtain a fused feature map. The first fused feature map is then downsampled layer by layer from bottom to top to half of the original size, and element-wise added to the original feature map of the same size obtained through the 1*1 convolution jump link and the first fused feature map of the previous layer. The feature map of the three added features is passed through a 3*3 convolution layer to obtain a secondary fused feature map.

[0048] (2) The feature fusion network uses the Kaiming initialization method to initialize the network parameters and is trained simultaneously with other network components.

[0049] (3) Multi-scale feature maps are trained through a skip connection feature pyramid network to obtain three fused feature maps, whose feature map sizes are ([256, w / 4, h / 4], [256, w / 8, h / 8], [256, w / 16, h / 16], w, h are the original image sizes) from large to small, with the same number of channels but different feature map sizes. Feature maps are assigned to predict target ships of different sizes according to their sizes. The target with the maximum size of [0, 64] is predicted by the large-scale feature map; the target with the maximum size of [64, 128] is predicted by the medium-scale feature map; the target with the maximum size of [128, INF] is predicted by the small-scale feature map.

[0050] 3. Prediction based on regression model of key points

[0051] The key point-based regression model directly predicts from the three layers of fused feature maps of different sizes obtained in the previous step. The regression model can be divided into two parts: the key point confidence regression subnetwork and the key point distance regression subnetwork. The key point confidence regression subnetwork predicts the score map of the key point, and the position regression subnetwork predicts the distance from the key point to the four sides of the prediction box. The steps are as follows:

[0052] (1) The key point confidence regression subnetwork is composed of three 3x3 convolutional layers, one 1x1 convolutional layer and a sigmoid function in sequence. The score map output by the network has the same size as the input feature map. The value of the element on the score map represents the probability that there is a key point of the ship target at the corresponding position point in the original image. The structure of the regression subnetwork is consistent with that of the classification subnetwork. The number of channels of the output feature map is 4, representing the distance prediction value of the key point (t1, t2, t3, t4). The prediction box position is located by the key point position and the distance from the key point to the four sides of the prediction box. The distance between the key point and the prediction box edge is calculated by the distance prediction value and the learnable scale parameter. The formula is as follows:

[0053] d i =exp(l j )×exp(t i ), for i=1,2,3,4

[0054] Among them, d i is the distance prediction quantity, l j is the size parameter of the science department, t i is the network distance prediction quantity;

[0055] (2) The key point-based regression model uses the Kaiming initialization method to initialize the network parameters. The samples are first assigned to different feature maps for prediction according to the above multi-scale division rules, and then a confidence score map with the same size as the feature map is generated based on the sample center distance. The formula is as follows. The confidence regression subnetwork uses the cross-entropy loss function to calculate the category loss between the score map predicted by the network and the generated confidence score map. The position regression subnetwork first takes the points with a score greater than 0.5 on the confidence score map as key points, calculates the distance from the key points to the four sides of the sample box, and then uses the distance from the key points to the real sample and the distance from the network prediction to calculate the IoU loss function as the position loss of the model. The loss function includes the cross-entropy loss of the confidence regression subnetwork, the IoU loss of the position regression subnetwork, and the L1 regularization loss of the convolution kernel skeleton matrix in the strip convolution layer. Feature maps of different scales share a key point-based regression model, and three different learnable scale parameters are used in the position regression subnetwork to predict objects of different scales.

[0056]

[0057] Among them, l is the distance from the predicted point to the left width of the true label, r is the distance from the predicted point to the right width of the true label, t is the distance from the predicted point to the top width of the true label, and b is the distance from the predicted point to the bottom width of the true label; P center is the probability value of the confidence of the target prediction point.

[0058] (3) The model was trained using the stochastic gradient descent algorithm, with a learning rate of 0.001, which was reduced to one-tenth of the original value every 10,000 iterations, and a momentum parameter of 0.9. The model was trained for 90,000 rounds, with a batch size of 8.

[0059] (4) The three layers of fused feature maps of different scales are trained to generate confidence prediction score maps and position prediction matrices based on the key point regression model. The points with scores greater than 0.5 in the score map are selected as prediction key points, and the four-side distances are calculated using the position prediction matrix and calculation formula to finally obtain the prediction box position. Finally, the soft non-maximum suppression algorithm is used to exclude overlapping prediction boxes to obtain the final prediction result.

[0060] The method of the present invention realizes end-to-end pruning and training, improves the robustness and training speed of the model; designs a skip connection feature pyramid network to complete bottom-up and top-down fusion, so that feature maps of different scales have semantic information of different depths, and improves the inspection accuracy of targets of different scales; the regression model based on key points greatly reduces the number of candidate boxes, further reduces the number of parameters and computational complexity; the learnable scale parameters replace the prior box scale parameters, and improves the positioning performance of the model.

[0061] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A lightweight SAR image ship detection system based on strip pruning, characterized in that It includes a lightweight residual convolutional network based on strip convolution, a skip connection feature pyramid network and a regression model based on key points; the lightweight residual convolutional network based on strip convolution is used to extract feature maps of different depths, and the feature maps are passed into the skip connection feature pyramid network to obtain a fusion feature map that fully integrates deep features and shallow features, and the regression model based on key points is used to directly predict the score and prediction box position of each feature point; the lightweight residual convolutional network based on strip convolution is an improvement of ResNet50, specifically, ResNet50 is replaced by ResNet50. The standard 3*3 convolution kernel in the Net50 network is replaced by a strip convolution kernel and a convolution kernel skeleton matrix. The strip convolution layer is composed of a standard convolution layer and a convolution kernel skeleton matrix. During calculation, the convolution kernel is dot-multiplied with the convolution kernel skeleton matrix to form a general convolution kernel for convolution operation. The convolution kernel skeleton matrix is ​​composed of learnable parameters, each parameter corresponds to a strip vector on the convolution kernel. The larger the parameter, the greater the weight of the strip when extracting features. During the training process, when the parameter is lower than the threshold, the strip is removed to achieve strip pruning.

2. A lightweight SAR image ship detection system based on strip pruning according to claim 1, characterized in that The skip connection feature pyramid network fuses the extracted 4-layer feature maps twice from bottom to top and from top to bottom layer by layer to obtain 3-layer fused feature maps of different sizes.

3. The lightweight SAR image ship detection system based on strip pruning according to claim 1 is characterized in that The key point-based regression model is divided into two parts: a key point confidence regression subnetwork and a key point distance regression subnetwork. The key point confidence regression subnetwork predicts the score map of the key point, which is composed of three 3x3 convolutional layers, a 1x1 convolutional layer and a sigmoid function in sequence. The score map output by the network has the same size as the input feature map, and the value of the element on the score map indicates the probability that the key point of the ship target exists at the corresponding position point of the original image; the position regression subnetwork predicts the distance from the key point to the four sides of the prediction box, and the number of channels of the output feature map is 4, representing the distance prediction value (t1, t2, t3, t4) of the key point. The prediction box position is located by the key point position and the distance from the key point to the four sides of the prediction box. The distance between the key point and the prediction box edge is calculated by the distance prediction value and the learnable scale parameter. The formula is as follows: d i =exp(l j )×exp(t i ),for i=1,2,3,4 Among them, d i is the distance prediction quantity, l j is the size parameter of the science department, t i is the network distance prediction quantity.

4. A lightweight SAR image ship detection method based on strip pruning implemented by the system of claim 1, characterized in that Here are the steps: Step 1: Extract multi-scale feature maps using a lightweight residual convolutional network based on strip convolution The SAR image block is passed into the lightweight residual convolution network based on strip convolution for feature extraction. After 50 strip convolution layers, 4 feature maps with different depths and sizes are extracted. The sizes of the feature maps are [256, w / 4, h / 4], [512, w / 8, h / 8], [1024, w / 16, h / 16], [2048, w / 32, h / 32] from large to small, where w and h are the original image sizes. Step 2: Skip connection feature pyramid network to fuse different depth features The four feature maps of different sizes and depths obtained in step 1 are passed into the skip connection feature pyramid network to fuse the features of different layers. The skip connection feature pyramid network performs two feature fusions from bottom to top and top to bottom. The bottom-up fusion is similar to the feature pyramid network. The network first converts the number of channels of feature maps of different layers from 256, 512, 1024, and 2048 to 256 through a 1*1 convolution layer, and then the feature map is upsampled by 2 times from top to bottom through the nearest neighbor interpolation method and added to the feature map of the next layer element by element. The added feature map is then passed through a 3*3 convolution layer to obtain a fused feature map; the fused feature map is then passed through a bottom-to-bottom convolution layer. The feature maps are downsampled layer by layer to half of the original ones, and are element-wise added with the original feature maps of the same size obtained by 1*1 convolution jump link and the fused feature maps of the previous layer. The feature maps of the three added maps are then passed through a 3*3 convolution layer to obtain secondary fused feature maps [256, w / 4, h / 4], [256, w / 8, h / 8], [256, w / 16, h / 16], where w and h are the original image sizes. The feature maps are allocated to predict target ships of different sizes according to their sizes. The target with the maximum size of [0, 64] is predicted by the large-scale feature map; the target with the maximum size of [64, 128] is predicted by the medium-scale feature map. The maximum size of the object is [128, INF], which is predicted by the small-scale feature map; Step 3: Prediction based on keypoint regression model The 3-layer fusion feature maps of different scales are trained based on the key point regression model to produce the confidence prediction score map and position prediction matrix. The points with scores greater than 0.5 in the score map are selected as the prediction key points, and the four-side distance is calculated through the position prediction matrix and the calculation formula to finally get the prediction box position. Finally, the soft non-maximum suppression algorithm is used to exclude overlapping prediction boxes to get the final prediction result.

5. The method for detecting ships in lightweight SAR images based on strip pruning according to claim 4 is characterized in that The skip connection feature pyramid network uses the Kaiming initialization method to initialize the network parameters.

6. The lightweight SAR image ship detection method based on strip pruning according to claim 4 is characterized in that The key point-based regression model uses the Kaiming initialization method to initialize network parameters.

Images