SAR ship detection method based on new pyramid structure and mixed pool channel attention mechanism

Abstract

The application discloses a SAR ship detection method based on a novel pyramid structure and a mixed pooling channel attention mechanism, and comprises the following steps: constructing a SAR ship dataset, and dividing a training dataset and a test dataset; constructing a target detection model based on the novel pyramid structure and the mixed pooling channel attention mechanism; sending the training dataset into the constructed target detection model for training, and obtaining weight parameters in the target detection model; sending the test dataset into the trained target detection model, obtaining corresponding position information and categories, and completing SAR ship detection. The application can not only reduce system calculation complexity, but also improve target detection accuracy, provide more effective information for a target detection task, and realize accurate detection of a SAR ship target.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of SAR ship target detection, and particularly relates to a SAR ship detection method based on a new pyramid structure and a mixed pooling channel attention mechanism. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) has the imaging characteristics of all-weather, all-day, and not affected by weather, and has become one of the important means of earth observation. In SAR image interpretation, the target detection task has always been concerned, especially the detection of ship targets.

[0003] Traditional SAR ship detection methods require different sea clutter distribution models and need to adjust parameters according to different SAR images. In recent years, ship detection algorithms based on deep learning have gradually emerged. Lin et al. adopted a squeezing extraction mechanism to obtain high-quality multi-scale feature maps. Wei et al. proposed a high-resolution ship detection network, which fully utilized high-resolution and low-resolution convolution feature maps. Kang et al. fused deep semantic features and high-resolution features in different layers of feature maps, improving the detection performance of small ships. Ke et al. optimized multi-scale feature maps using Swin Transformer. Although Lin et al. proposed a multi-layer fusion algorithm to enrich the semantic information of shallow feature maps, it still cannot fully enrich shallow feature maps, resulting in low detection accuracy of small targets. At the same time, although the top feature maps contain rich semantic information, there is also a problem of information weakening after channel reduction. Furthermore, there is a lack of a suitable attention mechanism to highlight the key features of ships, resulting in high false alarms in detection.

[0004] Therefore, a new technical solution is needed to solve these problems. SUMMARY

[0005] The purpose of the application is to overcome the deficiencies in the prior art, provide a SAR ship detection method based on a new pyramid structure and a mixed pooling channel attention mechanism, which can not only reduce system computational complexity, but also improve target detection accuracy, provide more effective information for target detection tasks, and realize accurate detection of SAR ship targets.

[0006] Technical scheme: To achieve the above purpose, the application provides a SAR ship detection method based on a new pyramid structure and a mixed pooling channel attention mechanism, comprising the following steps:

[0007] S1: Construct a SAR ship dataset and divide it into a training dataset and a test dataset;

[0008] S2: Based on the new pyramid structure and the mixed pooling channel attention mechanism, construct a target detection model;

[0009] S3: Put the training data set into the built target detection model for training to obtain the weight parameters in the target detection model;

[0010] S4: Put the test data set into the trained target detection model to obtain the corresponding position information and category, and complete the SAR ship detection.

[0011] Further, the step S1 of preprocessing the built SAR ship data set includes horizontal flip and size unification and other preprocessing operations.

[0012] The detection process of the target detection model in the step S2 is as follows:

[0013] A1: Feature detection is performed using Convnext to obtain corresponding multi-scale feature mapping;

[0014] A2: The feature mapping obtained in A1 is sent to a mixed pooling channel attention mechanism module to highlight important feature mappings of the SAR ship and obtain a feature map;

[0015] A3: The feature map obtained in A2 is sent to a new pyramid structure for enriching semantic information of feature maps of each layer;

[0016] A4: All feature maps obtained in A3 are sent to RPN to obtain candidate boxes;

[0017] A5: The candidate boxes obtained in A4 are sent to ROI POOLING to unify and standardize the size;

[0018] A6: The candidate boxes obtained in A5 are sent to a double-branch prediction structure consistent with Faster R-CNN to obtain corresponding category and position prediction information. The category loss function and the position regression loss function are respectively represented by L1Loss and CrossEntropyLoss, and the formulas are as follows:

[0019]

[0020]

[0021] Wherein, n represents the number of samples, y i represents the true value, represents the predicted value.

[0022] Further, the Convnext in the step A1 is a pure convolutional network, and the Convnext as a feature extraction structure has the following 5 parts in turn:

[0023] 1 A non-overlapping convolution with a convolution kernel size of 4 and a step size of 4 and a layer normalization structure;

[0024] 3 consecutive Convnext blocks with an output channel of 96;

[0025] 1 down-sampling layer and 3 consecutive Convnext blocks with an output channel of 192;

[0026] 1 down-sampling layer and 9 consecutive Convnext blocks with an output channel of 384;

[0027] 1 down-sampling layer and 3 consecutive Convnext blocks with an output channel of 768.

[0028] The Convnext block in the Convnext is composed of the following three parts in turn:

[0029] A deep convolution structure with a convolution kernel size of 7, a step of 1, and padding of 3, and a layer normalization structure;

[0030] A convolution structure with a convolution kernel size of 1 and a step of 1, and a GELU activation function;

[0031] A convolution structure with a convolution kernel size of 1 and a step of 1, a Layer Scale, and a Drop Path.

[0032] Further, the running process of the mixed channel attention mechanism module in step A2 is as follows:

[0033] B1: The feature map is subjected to three different channel dimension pooling, namely maximum pooling, average pooling and soft pooling, to obtain three one-dimensional vectors, and the number of elements of the three different vectors is consistent with the number of channels of the input feature map;

[0034] B2: The three different one-dimensional vectors are respectively subjected to three fully connected layers to obtain three one-dimensional vectors again, wherein the number of nodes of the second fully connected layer is reduced by 7 times compared with the first layer to reduce the calculation cost;

[0035] B3: The three one-dimensional vectors obtained in B2 are added to obtain a one-dimensional vector;

[0036] B4: The one-dimensional vector obtained in B3 is processed using a sigmoid activation function to obtain a final weight vector.

[0037] Further, the design process of the new pyramid structure in step A3 is as follows:

[0038] C1: A multi-scale feature structure (from top to bottom, F4, F3, F2, F1) is known, and the number of channels of the multi-scale feature structure from top to bottom is (768, 384, 192, 96), and the size decreases by 1 / 2 from top to bottom;

[0039] C2: Use a 1*1 convolutional structure to change the number of channels in the original feature structure, transforming the original number of channels (768,384,192,96) into (256,256,256,256);

[0040] C3: Fi is multiplied by 2 and added to Fi-1 according to the principle of adding corresponding elements. Then, a 3*3 convolutional structure is used to remove the confusion after addition. The lowest feature map after the operation is selected as the lowest layer of the new feature pyramid structure.

[0041] C4: Select the top three feature maps after C3 operation and repeat C3 operation. Select the bottom feature map after the operation as the second layer of the new pyramid structure.

[0042] C5: Select the top three feature maps after C4 operation and repeat C3 operation. Select the bottom feature map after the operation as the third layer of the new pyramid structure.

[0043] C6: After feature layer F4 is processed by a top-level feature enhancement module, it is summed with the top-level feature map processed by C3 to obtain the top-level feature map of the new pyramid structure.

[0044] C7: The above steps result in a complete new pyramid structure.

[0045] The operation flow of the top-level feature enhancement module in C6 is as follows:

[0046] D1: Reconstruct the top-level feature map according to three scales (0.1*w*h, 0.2*w*h, 0.3*w*h);

[0047] D2: Enlarge the feature maps of three different scales to their original size using the nearest neighbor insertion algorithm;

[0048] D3: To mitigate the confusion caused by the nearest neighbor interpolation method in D2, a spatial attention mechanism is adopted to obtain three different spatial weights. The three different spatial weights are applied to the corresponding spatial matrix graphs to obtain three spatial feature maps after removing interpolation confusion.

[0049] D4: Add the corresponding elements of the three different spatial feature maps to obtain semantic information that combines the three different spatial features.

[0050] Furthermore, in step A6, category and location prediction information are obtained through a category loss function and a location regression loss function. The category loss function and the location regression loss function are represented by L1Loss and CrossEntropyLoss, respectively, and the formulas are shown below:

[0051]

[0052]

[0053] wherein n represents the number of samples, y i represents the true value, represents the predicted value.

[0054] The Convnext convolution extraction structure, the mixed pooling channel attention mechanism module and the novel pyramid structure used in the application are innovations of the application, which are different from the prior art. The application can effectively solve the problems of low feature mapping quality, high false alarm rate, insufficient semantic information of shallow layer features, damage of top layer high quality information due to channel reduction and the like, and improve the accuracy of SAR ship target detection.

[0055] The core of the method of the application is to detect the SAR ship through a target detection model. The target detection model mainly comprises a novel pyramid structure and a mixed pooling channel attention mechanism module. The novel pyramid structure mainly comprises two parts: the first part enriches the semantic information of the shallow layer feature map through certain feature enhancement means. The second part adopts three different spatial semantic information to make up for the semantic information loss of the top layer feature map in order to avoid the semantic information loss of the top layer feature map after channel reduction. The mixed pooling channel attention mechanism module highlights the effective feature mapping and suppresses the redundant feature mapping by combining the advantages of three different poolings, considering the relationship between different channels and the feature attributes of each channel. The method of the application can effectively improve the detection accuracy of the SAR ship target.

[0056] Advantages: Compared with the prior art, the application has the following advantages:

[0057] 1. Compared with the existing extraction structure, Convenxt-T is a pure convolutional neural network which imitates the Swin Transformer to adjust parameters while ensuring translational invariance. Therefore, the network combines the advantages of pure convolutional neural networks and attention mechanism networks, and can extract relatively high quality feature mapping in a complex marine background.

[0058] 2. Compared with the existing attention mechanism module, the mixed pooling channel attention mechanism module is a channel attention mechanism which combines the advantages of maximum pooling, average pooling and soft pooling, aims to generate corresponding weights for each channel matrix mapping, highlights the SAR ship target feature information and suppresses irrelevant feature information, thereby reducing the false alarm rate of detection.

[0059] 3、Compared with the existing feature enhancement method, the new pyramid structure fully enriches the semantic information of the shallow feature through the top-down and left-to-right information fusion mode. The top feature map has rich semantic information, but due to the reduction of the channel, the top high-quality semantic information is damaged, in order to solve this problem, the application uses a top feature enhancement module to solve this problem, thereby improving the accuracy of SAR ship target detection. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of the method of the application;

[0061] Figure 2 is a schematic diagram of the target detection model in the application;

[0062] Figure 3 is a Convnext block structure diagram;

[0063] Figure 4 is a new pyramid structure diagram;

[0064] Figure 5 is a top feature enhancement module diagram in the new pyramid structure;

[0065] Figure 6 is a mixed pooling channel attention mechanism structure diagram;

[0066] Figure 7 is an ablation experiment result diagram about the new pyramid structure (FEFPN);

[0067] Figure 8 is an ablation experiment result diagram about whether to use the top feature enhancement module (TLFE);

[0068] Figure 9 is an ablation experiment result diagram about the mixed pooling channel attention mechanism module (MPCA);

[0069] Figure 10 is a comparison experiment data diagram of the method of the application and other methods;

[0070] Figure 11 is a visualization result of four detectors for nearshore images;

[0071] Figure 12 is a visualization result of four detectors for farshore images;

[0072] Figure 13 is an effect diagram of three attention mechanism modules for ship identification, wherein MPCA is a mixed pooling channel attention mechanism module. DETAILED DESCRIPTION

[0073] The present application will be further illustrated below in conjunction with the accompanying drawings and specific embodiments, it should be understood that these embodiments are only used to illustrate the present application and are not used to limit the scope of the present application, and after reading the present application, various equivalent modifications of the present application by those skilled in the art fall within the scope defined by the appended claims.

[0074] As shown in Figure 1 The present application provides a SAR ship detection method based on a new pyramid structure and a mixed pooling channel attention mechanism, comprising the following steps:

[0075] S1: Constructing a SAR ship dataset:

[0076] In this embodiment, the SSDD ship dataset is used, and the dataset will be preprocessed during the experiment, the horizontal flip is 0.5 times, and the size is unified to 608*608; the preprocessed dataset is divided into training set and test set in the ratio of 8:2;

[0077] S2: Constructing a target detection model based on a new pyramid structure and a mixed pooling channel attention mechanism;

[0078] S3: Sending the training dataset into the constructed target detection model for training to obtain the weight parameters in the target detection model, and obtaining the trained target detection model;

[0079] S4: Sending the test dataset into the trained target detection model to obtain the corresponding position information and category, and completing the SAR ship detection.

[0080] As shown in Figure 2 The detection process of the target detection model in step S2 is as follows:

[0081] A1: Using Convnext for feature detection to obtain corresponding multi-scale feature mapping;

[0082] A2: Sending the feature mapping obtained in A1 into the mixed pooling channel attention mechanism module to highlight the important feature mapping of the SAR ship and obtain the feature map;

[0083] A3: Sending the feature map obtained in A2 into the new pyramid structure to enrich the semantic information of the feature maps of each layer;

[0084] A4: Sending all the feature maps obtained in A3 into RPN to obtain the candidate frame;

[0085] A5: Sending the candidate frame obtained in A4 into ROI POOLING to unify the size;

[0086] A6: The candidate frame obtained in A5 is sent to a double-branch prediction structure consistent with Faster R-CNN to obtain corresponding class and position prediction information.

[0087] In step A1, Convnext is a pure convolutional network. The Convnext as a feature extraction structure is composed of the following 5 parts in turn:

[0088] 1 convolution kernel size 4, step 4, no overlap convolution and a layer normalization structure;

[0089] 3 consecutive Convnext structure blocks, output channel 96;

[0090] 1 down-sampling layer and 3 consecutive Convnext structure blocks, output channel 192;

[0091] 1 down-sampling layer and 9 consecutive Convnext structure blocks, output channel 384;

[0092] 1 down-sampling layer and 3 consecutive Convnext structure blocks, output channel 768.

[0093] As shown in Figure 3 , the Convnext structure block in Convnext is composed of the following 3 parts in turn:

[0094] 1 convolution kernel size 7, step 1, padding 3, depth convolution structure and a layer normalization structure;

[0095] 1 convolution kernel size 1, step 1, convolution structure and a GELU activation function;

[0096] 1 convolution kernel size 1, step 1, convolution structure, a Layer Scale and a Drop Path.

[0097] As shown in Figure 6 , the running process of the mixed pooling channel attention mechanism module in step A2 is as follows:

[0098] B1: The feature map is subjected to three different poolings in the channel dimension, namely maximum pooling, average pooling and soft pooling, to obtain three one-dimensional vectors. The number of elements of the three different vectors is consistent with the number of channels of the input feature map;

[0099] B2: The three different one-dimensional vectors are respectively subjected to three fully connected layers to obtain three one-dimensional vectors again. The number of nodes of the second fully connected layer is reduced by 7 times compared with the first layer to reduce the calculation cost;

[0100] B3: The three one-dimensional vectors obtained in B2 are summed to obtain a one-dimensional vector;

[0101] B4: using a sigmoid activation function to process the one-dimensional vector obtained in B3, and obtaining the final weight vector.

[0102] As shown in Figure 4 , the design process of the new pyramid structure in step A3 is as follows:

[0103] C1: a multi-scale feature structure (from top to bottom, F4, F3, F2, F1) is known, and the number of channels from top to bottom is (768, 384, 192, 96), and the size decreases by 1 / 2 from top to bottom;

[0104] C2: a 1*1 convolution structure is used to change the channel number of the original feature structure, and the original channel number (768, 384, 192, 96) is transformed into (256, 256, 256, 256);

[0105] C3: expand Fi by 2 times, and add Fi-1 according to the principle of corresponding element addition, and then remove the confusion after addition through a 3*3 convolution structure, and select the bottommost feature map after the operation as the bottommost layer of the new feature pyramid structure;

[0106] C4: select the top three layers of feature maps after C3 operation and repeat C3 operation, and select the bottommost feature map after the operation as the second layer of the new pyramid structure;

[0107] C5: select the top three layers of feature maps after C4 operation and repeat C3 operation, and select the bottommost feature map after the operation as the third layer of the new pyramid structure;

[0108] C6: the top layer feature enhancement module processes the feature layer F4, and performs corresponding element addition operation with the top layer feature map after C3 processing, to obtain the top layer feature map of the new pyramid structure;

[0109] C7: the complete new pyramid structure is obtained through the above.

[0110] As shown in Figure 5 , the operation process of the top layer feature enhancement module in C6 is as follows:

[0111] D1: reconstruct the size of the top layer feature map according to three proportions (0.1*w*h, 0.2*w*h, 0.3*w*h);

[0112] D2: expand the feature maps of three different proportions to the original size through the adjacent insertion algorithm;

[0113] D3: In order to alleviate the confusion caused by the adjacent interpolation method in D2, a spatial attention mechanism is adopted to obtain three different spatial weights, and the three different spatial weights are applied to the corresponding spatial matrix diagram to obtain three spatial feature maps after removing the interpolation confusion;

[0114] D4: The three different spatial feature maps are subjected to corresponding element addition operation to obtain semantic information combined with three different spatial features.

[0115] The category and position prediction information in step A6 is obtained by a category loss function and a position regression loss function, and the category loss function and the position regression loss function are represented by L1Loss and CrossEntropyLoss respectively, and the formulas are as follows:

[0116]

[0117]

[0118] Wherein, n represents the number of samples, y i represents the true value, represents the predicted value.

[0119] Based on the above scheme, in order to verify the effectiveness of the method of the application, the embodiment verifies the effect of the application by experiment, as follows:

[0120] The ablation experiment on the new pyramid structure (FEFPN) is carried out, as shown in Figure 7 The experimental results of the ablation experiment are shown, and the experimental data shows that Convnext and the new pyramid structure (FEFPN) are better than other advanced structures.

[0121] The ablation experiment on whether to use the top layer feature enhancement module (TLFE) is carried out, as shown in Figure 8 The experimental results of the ablation experiment are shown, and the experimental data shows that the effect of using the top layer feature enhancement module (TLFE) is more significant.

[0122] The ablation experiment on the mixed pooling channel attention mechanism module (MPCA) is carried out, as shown in Figure 9 The experimental results of the ablation experiment are shown, and the experimental data shows that the mixed pooling channel attention mechanism module (MPCA) is better than other advanced attention mechanism modules.

[0123] As shown in Figure 10 The experimental results of the comparative experiment are shown, and the experimental data shows that the method of the application is better than other advanced methods.

[0124] As shown in Figure 11As shown, the visualization results of four methods on near-shore images are presented, and the visualization results show that the method of the present invention has the best effect.

[0125] like Figure 12 As shown, the visualization results of four methods on images of distant coastlines are presented, and the visualization results show that the method of the present invention has the best effect.

[0126] like Figure 13 As shown, the visualization results of three attention mechanism modules on the experimental dataset are presented. The visualization results show that the method of the present invention has the best effect.

[0127] In summary, it is evident that the present invention has the following advantages compared to the prior art:

[0128] 1. Compared to existing extraction structures, Convenxt-T is a pure convolutional neural network that, while maintaining its translation invariance, mimics the Swing Transformer to adjust parameters. Therefore, this network combines the advantages of pure convolutional neural networks and attention-based networks, enabling it to extract relatively high-quality feature maps in complex marine environments.

[0129] 2. Compared to existing attention mechanism modules, the hybrid pooling channel attention mechanism module is a channel attention mechanism that combines the advantages of max pooling, average pooling, and soft pooling. It aims to generate corresponding weights for each channel matrix mapping, highlighting the SAR ship target feature information and suppressing irrelevant feature information, thereby reducing the false alarm rate.

[0130] 3. Compared to existing feature enhancement methods, the novel pyramid structure enriches the semantic information of shallow features through top-down and left-to-right information fusion. While the top-level feature map possesses rich semantic information, the high-quality semantic information is compromised due to channel reduction. To address this issue, this invention employs a top-level feature enhancement module, thereby improving the accuracy of SAR ship target detection.

Claims

1. A SAR ship detection method based on a novel pyramid structure and a hybrid pooling channel attention mechanism, characterized in that, Includes the following steps: S1: Construct a SAR ship dataset and divide it into training and test datasets; S2: Construct an object detection model based on a novel pyramid structure and a hybrid pooling channel attention mechanism; S3: Feed the training dataset into the constructed object detection model for training, and obtain the weight parameters in the object detection model; S4: Input the test dataset into the trained target detection model to obtain the corresponding location information and category, and complete the SAR ship detection; The detection process of the target detection model in step S2 is as follows: A1: Use Convnext for feature detection to obtain the corresponding multi-scale feature maps; A2: The feature map obtained from A1 is fed into the hybrid pooling channel attention mechanism module to highlight the important feature maps of SAR ships and obtain feature maps; A3: The feature map obtained from A2 is fed into the new pyramid structure to enrich the semantic information of the feature maps at each layer; A4: Input all feature maps obtained from A3 into the RPN and obtain candidate boxes; A5: Send the candidate boxes obtained from A4 into ROIPOOLING to standardize their dimensions; A6: Feed the candidate boxes obtained from A5 into a two-branch prediction structure consistent with Faster R-CNN to obtain the corresponding category and location prediction information.

2. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 1, characterized in that, In step S1, the constructed SAR ship dataset is preprocessed, including horizontal flipping and size unification.

3. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 1, characterized in that, In step A6, category and location prediction information are obtained through the category loss function and the location regression loss function. The category loss function and the location regression loss function are represented by L1Loss and CrossEntropyLoss, respectively, and the formulas are shown below: Where n represents the number of samples, y i Represents the true value. This represents the predicted value.

4. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 1, characterized in that, In step A1, Convnext is a pure convolutional network. As a feature extraction structure, Convnext consists of the following five parts in sequence: One uncovered convolutional layer with a kernel size of 4 and a stride of 4, and a layer normalization structure; Three consecutive Convnext structure blocks, with 96 output channels; One downsampling layer and three consecutive Convnext structure blocks, with 192 output channels; One downsampling layer and nine consecutive Convnext blocks, with 384 output channels; It has one downsampling layer and three consecutive Convnext structure blocks, with 768 output channels.

5. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 4, characterized in that, The Convnext structure block in the Convnext is composed of the following three parts in sequence: A depthwise convolutional structure with a kernel size of 7, a stride of 1, and padding of 3, and a layer normalization structure; A convolutional structure with a kernel size of 1 and a stride of 1, and a GELU activation function; A convolutional structure with a kernel size of 1 and a stride of 1, a Layer Scale, and a Drop Path.

6. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 1, characterized in that, The operation flow of the hybrid pooling channel attention mechanism module in step A2 is as follows: B1: The feature map undergoes three different pooling methods along the channel dimension: max pooling, average pooling, and soft pooling, resulting in three one-dimensional vectors. The number of elements in the three different vectors is the same as the number of channels in the input feature map. B2: Three different one-dimensional vectors are passed through three fully connected layers to obtain three more one-dimensional vectors. The number of nodes in the second fully connected layer is reduced by 7 times compared to the first layer to reduce computational cost. B3: Sum the three one-dimensional vectors obtained in B2 to get a single one-dimensional vector; B4: The one-dimensional vector obtained from B3 is processed using the sigmoid activation function to obtain the final weight vector.

7. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 1, characterized in that, The design process for the novel pyramid structure in step A3 is as follows: C1: Given a multi-scale feature structure, from top to bottom, the numbers are F4, F3, F2, and F1. The number of channels in this multi-scale feature structure from top to bottom is (768, 384, 192, 96), and the size decreases by 1 / 2 from top to bottom. C2: Use a 1*1 convolutional structure to change the number of channels in the original feature structure, transforming the original number of channels (768,384,192,96) into (256,256,256,256); C3: Fi is multiplied by 2 and added to Fi-1 according to the principle of adding corresponding elements. Then, a 3*3 convolutional structure is used to remove the confusion after addition. The lowest feature map after the operation is selected as the lowest layer of the new feature pyramid structure. C4: Select the top three feature maps after C3 operation and repeat C3 operation. Select the bottom feature map after the operation as the second layer of the new pyramid structure. C5: Select the top three feature maps after C4 operation and repeat C3 operation. Select the bottom feature map after the operation as the third layer of the new pyramid structure. C6: After feature layer F4 is processed by a top-level feature enhancement module, it is summed with the top-level feature map processed by C3 to obtain the top-level feature map of the new pyramid structure. C7: The above steps result in a complete new pyramid structure.

8. The SAR ship detection method based on a novel pyramid structure and hybrid pooling channel attention mechanism according to claim 7, characterized in that, The operation flow of the top-level feature enhancement module in C6 is as follows: D1: Reconstruct the size of the top-level feature map according to three proportions; D2: Enlarge the feature maps of three different scales to their original size using the nearest neighbor insertion algorithm; D3: A spatial attention mechanism is adopted to obtain three different spatial weights. The three different spatial weights are applied to the corresponding spatial matrix graphs to obtain three spatial feature maps after removing interpolation confusion. D4: Add the corresponding elements of the three different spatial feature maps to obtain semantic information that combines the three different spatial features.

Images