Yangtze river surface ship identification method based on MFD-LRASPP network
By constructing the MFD-LRASPP network, the problem of low ship segmentation accuracy under occlusion conditions is solved, efficient ship recognition in complex ocean scenes is achieved, the feature extraction and discrimination capabilities of the model are enhanced, and the segmentation accuracy is improved.
Patent Information
- Application Number
- CN202510661838.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-22
- Publication Date
- 2025-09-23
AI Technical Summary
Existing technologies have low accuracy in ship segmentation under occlusion, and the lack of effective datasets limits model training and evaluation, making it difficult to achieve accurate ship instance segmentation in complex ocean scenes.
A network based on MFD-LRASPP is constructed, including an inverted residual module, a lightweight multi-scale feature extraction module and a bottleneck feature discrimination module. Through deep separable convolution and residual structure, the feature extraction capability and discrimination accuracy are improved. A self-built Yangtze River ship dataset is also built to enhance the robustness of the model.
The accuracy and robustness of ship segmentation are improved, and the ship boundaries can be accurately identified in complex occlusion scenes, which enhances the generalization ability and computational efficiency of the model.
Smart Images
Figure CN120689774A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of computer vision image processing and surface vessel recognition, and relates to a method for recognizing vessels on the Yangtze River based on an MFD-LRASPP network. Background Art
[0002] In ports and offshore areas, high-density ship traffic can lead to dangerous situations, necessitating reliable traffic management and safety control measures. Automatic vessel identification is becoming increasingly important for maritime surveillance, monitoring, traffic supervision, and the regulation of illegal fishing and cargo transit. In this context, ship detection and segmentation technologies are crucial for obtaining more precise position and shape information for ship identification.
[0003] Ship imagery primarily includes synthetic aperture radar (SAR) images, infrared images, and visible light images. SAR-based methods offer the advantage of relatively mature all-weather monitoring capabilities. However, SAR images are often subject to significant gamma-distributed speckle noise, which can make identifying small edges challenging, and their spectral information is not conducive to ship segmentation. In contrast, infrared images have good penetration but are less sensitive to ambient brightness. Their resolution and signal-to-noise ratio are low, and target color and texture information is limited, affecting segmentation accuracy. Compared to these two types of imagery, visible light images contain more detailed feature information and offer higher resolution.
[0004] Traditional image segmentation methods, such as thresholding, edge detection, and Gaussian mixture models, were widely used in the early days of ship segmentation. For example, Li et al. used enhanced thresholding and region growing techniques to achieve ship segmentation in images. Lee et al. proposed an improved GrabCut method that uses clustering technology to reduce image noise and improve the accuracy of ship segmentation in SAR images. Ren et al. used edge detection technology to quickly identify edge information in images and connected the pixels of ship targets through region growing, achieving good ship instance segmentation. However, traditional instance segmentation methods suffer from poor feature generalization and low detection accuracy. Compared with traditional segmentation methods, convolutional neural networks (CNNs) provide more accurate segmentation accuracy and higher algorithm generalization capabilities. Currently, popular models that use masked data include SOLOv2, BoxInst, BlendMask, and RefineMask, all of which have achieved many fruitful results.
[0005] While the CNN-based ship segmentation algorithms mentioned above are generally effective, few algorithms can effectively address the problem of ship segmentation under occlusion. This is because occlusion can partially or completely hide the ship's structure, making it difficult for segmentation algorithms to accurately identify the ship's boundaries. Occlusion results in partial or complete loss of structural information about the ship target, severely interfering with boundary extraction. Common occlusion scenarios include overlapping ships, small ships obscured by larger targets, and ship visibility hampered by environmental factors such as waves or weather. Occlusion divides the target into several segments, resulting in instance fragmentation. These occlusion phenomena hinder the extraction of key features such as the ship's outline, shape, and texture, which in turn affects the accuracy of ship instance segmentation in complex ocean scenes.
[0006] In addition, there are relatively few publicly available visible light ship segmentation datasets, which further limits the training and evaluation of the model. There is an urgent need for exploration and breakthroughs in data acquisition and network structure improvement. Summary of the Invention
[0007] Purpose of the invention: The purpose of the present invention is to provide a method for identifying ships on the Yangtze River based on the MFD-LRASPP network, which not only enhances the multi-scale feature extraction capability of the network, but also performs discriminative learning on the obtained bottleneck information, thereby improving the classification capability of the segmentation network at the pixel level.
[0008] Technical solution: The method for identifying ships on the Yangtze River based on the MFD-LRASPP network of the present invention comprises the following steps:
[0009] (1) Segment the pre-filmed Yangtze River ship videos and create a dataset, and then preprocess the obtained Yangtze River ship dataset;
[0010] (2) Construct a vessel segmentation network based on the MFD-LRASPP model;
[0011] (3) Train the constructed ship segmentation network based on the MFD-LRASPP model and adjust the training parameters of the model multiple times until the optimal network model weight is obtained;
[0012] (4) Verify and evaluate the performance of the obtained optimal network model weights;
[0013] (5) The optimal network model weights after verification and evaluation are loaded, and the real surface ship navigation images are predicted to finally obtain the ship segmentation results.
[0014] Furthermore, the data set in step (1) is divided into a training set and a validation set in a ratio of 8:2; and the preprocessing includes normalization and scaling.
[0015] Furthermore, the implementation process of pre-shooting the video of the ships on the Yangtze River in step (1) is:
[0016] First, the video was captured using a drone placed over the Yangtze River's waterways. The video parameters were a resolution of 3840×2160 and a frame rate of 60, with each waterway divided into sections every 200 meters.
[0017] Secondly, we captured images at 1-second intervals from the video, filtered out images with small content differences, and labeled the ships in the images using the labelme annotation tool.
[0018] Finally, the obtained json annotation file is converted into a segmentation label mask image, and the dataset is divided into a training set and a validation set.
[0019] Furthermore, the ship segmentation network in step (2) includes an inverted residual module, a lightweight multi-scale feature extraction module, a bottleneck feature discrimination module and a segmentation head LRASPP Head.
[0020] Furthermore, the inverted residual module adopts depth-wise separable convolution, including channel-by-channel convolution and point-by-point convolution;
[0021] The number of parameters P of standard convolution s and computational effort F s as follows:
[0022] P s =K 2 ×M×N (1)
[0023] F s =H out ×W out ×M×K 2 ×N (2)
[0024] The number of parameters P of depthwise separable convolution DP and computational effort F DP as follows:
[0025] P DP =K 2 ×M+M×N (3)
[0026] F DP =K 2 ×H out ×W out ×M+H out ×W out ×M×N (4)
[0027] Therefore, the ratio of the number of parameters P and the number of computations F between standard convolution and depthwise separable convolution are as follows:
[0028]
[0029] Where H out and W out They are the height and width of the output feature map, M is the number of channels of the input feature map, K is the convolution kernel size, and N is the number of convolution kernels, which is also the number of channels of the output feature map.
[0030] Furthermore, the lightweight multi-scale feature extraction module is: first, a 1×1 convolution is used to map the feature map after the fully connected layer from the low-dimensional space to the high-dimensional space; then, two feature branches with convolution scales of 3×3 and 5×5 are used, and finally, splicing is performed on the channel to obtain multi-scale features.
[0031] Furthermore, the bottleneck feature discrimination module is: first, a layer normalization operation is performed on the bottleneck feature, and then a partial convolution is performed; and then a second feature discrimination is performed through a gating mechanism controlled by a depth-wise separable convolution.
[0032] Furthermore, the implementation process of the single training of the ship segmentation network based on the MFD-LRASPP model in step (3) is as follows: the features after data enhancement are used as the input of the convolutional network to perform a convolution operation with kernel_size=1, and then the feature vector S1 is obtained by performing a Relu6 activation function and layer normalization;
[0033] First, in the backbone network, input data S1 is processed through a 3×3 convolution layer, batch normalization, and layer normalization to generate S11. S11 is then activated through adaptive average pooling, 1×1 convolution, ReLU, and Hard Sigmoid to generate S12. The dot product of the two is then processed through a 1×1 convolution to form S2, which is then added to S1 to obtain the feature vector S3. S3 is processed through multiple inverted residual modules to generate S4 and S5. In the lightweight multi-scale feature extraction module, S5 extracts features through three different convolution branches and concatenates them into S6, which is then processed through two layers of LMFB to obtain S7 and S8.
[0034] Next, S9 and S10 are generated in sequence through the inverted residual module, and then S11 and S12 are obtained through two layers of LMFB. S12 passes through multiple inverted residual modules with different expansion rates to finally generate S14-S17.
[0035] Then, in the bottleneck layer, S17 obtains the bottleneck feature S18 through 1×1 convolution, and is optimized in the bottleneck feature discrimination module to generate S181 and add it to S18 to form the final feature S19, which is sent to the LR-ASPP Head for image segmentation.
[0036] Furthermore, the performance of the optimal network model weight obtained in step (4) is verified and evaluated as follows: the Yangtze River surface vessel recognition network based on the constructed MFD-LRASPP network is trained, and the network model is iteratively trained multiple times, and the training parameters of the model are adjusted: learning rate, batch size, number of iterations and optimizer, until the optimal model parameters are obtained;
[0037] The best training model weight is used as the pre-training model to verify the data images in the validation set, and the performance of the model is evaluated by four indicators: accuracy, precision, recall rate and mAP;
[0038] The calculation formulas for each indicator are as follows:
[0039]
[0040] In the formula, TP represents true positive examples, FP represents false positive examples, FN represents false negative examples, TN represents true negative examples, C represents the total number of categories, AP represents the total number of categories, and i represents the AP value of the i-th class.
[0041] Furthermore, the implementation process of predicting the surface ship navigation image in step (5) is: first, loading the optimized network model weights to ensure the best prediction accuracy; then, inputting the image to be processed into the deep learning model, which analyzes the image based on its learned features; finally, the model outputs the ship segmentation result, that is, accurately identifying and segmenting the ship area in the image.
[0042] Beneficial effects: Compared with the prior art, the present invention has the following significant features: 1. The present invention has built a dataset of ships on the Yangtze River. On the one hand, the dataset contains scenes under various lighting conditions and ships at perspective distances, which is beneficial to improving the robustness of the model. On the other hand, the dataset taken from the angle of a drone is helpful for deploying flight equipment for monitoring the Yangtze River surface; 2. The present invention proposes a bottleneck feature discrimination module FDB, which is a two-layer gating mechanism composed of partial convolution and depthwise separable convolution respectively, and performs two feature discriminations on bottleneck features; while discriminating local features and improving model accuracy, it reduces the number of parameters and calculations, and adopts a residual structure to prevent the network from generating gradient explosion problems; 3. The present invention proposes a lightweight multi-scale feature extraction module LMFB to aggregate global features. This module not only effectively expands the receptive field of the neural network by synchronously realizing information interaction in the spatial and channel dimensions, but also significantly enhances the feature extraction capability and parameter utilization efficiency; by introducing a multi-branch channel fusion structure, while maintaining model accuracy, the deep learning model has a more compact architectural design while maintaining excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is an operational flow chart of the present invention;
[0044] Figure 2 This is a schematic diagram of the network structure of the LMFB module in the present invention;
[0045] Figure 3 This is a schematic diagram of the network structure of the FDB module in the present invention;
[0046] Figure 4 Schematic diagram of the network structure of the Inverted Residual module used in the present invention;
[0047] Figure 5 This is a schematic diagram of the network structure of the Yangtze River surface vessel identification method based on the MFD-LRASPP network in the present invention;
[0048] Figure 6 It is a training fitting process diagram of the present invention. DETAILED DESCRIPTION
[0049] The specific technical solutions of the present invention are further described in detail below with reference to specific examples.
[0050] This embodiment involves a large number of variables, and each variable is described below, as shown in Table 1.
[0051] Table 1 Variable description table
[0052]
[0053]
[0054] As shown in the figure, the method for identifying ships on the Yangtze River based on the MFD-LRASPP network of the present invention is characterized by comprising the following steps:
[0055] (1) The pre-shot Yangtze River ship videos were segmented and made into a dataset, which was divided into a training set and a test set in a ratio of 8:2. The obtained Yangtze River ship dataset was then preprocessed, including normalization and scaling.
[0056] (2) Construct a vessel segmentation network based on the MFD-LRASPP model;
[0057] (3) Train the vessel segmentation network of the constructed MFD-LRASPP model and adjust the training parameters of the model multiple times until the optimal network model weight is obtained.
[0058] (4) Verify and evaluate the optimal network model weight performance;
[0059] (5) The optimal network model weights that have been verified and evaluated are loaded, and the real surface ship navigation images are predicted to obtain the ship segmentation results.
[0060] The implementation process of pre-shooting the video of the ships on the Yangtze River in the step (1) is as follows: first, the video is cruised on the Dantu straight waterway, the Jiaoshan waterway, the Yangtze River estuary straight waterway, and the Yizheng waterway of the Yangtze River using a DJI AVATA 2 drone, wherein the video parameters are a resolution of 3840×2160 and a frame rate of 60, and each waterway is divided into a river section every 200 meters; second, the video is intercepted at an interval of 1 second, and the image materials with small content differences are screened out, and the ships in the image are annotated using the labelme annotation tool; finally, the obtained json annotation file is converted into a segmentation label mask (PNG format) image, and the data set is divided into a training set and a test set.
[0061] The vessel segmentation network in step (2) includes a basic feature extraction module Inverted Residual Block, a lightweight multi-scale feature extraction module LMFB, a bottleneck feature discrimination module (FDB) and a segmentation head LRASPP Head.
[0062] The Inverted Residual Block adopts depth-separable convolution, including channel-by-channel convolution and point-by-point convolution;
[0063] The number of parameters P of standard convolution s and computational effort F s as follows:
[0064] P s =K 2 ×M×N (1)
[0065] F s =H out ×W out ×M×K 2 ×N (2)
[0066] The number of parameters P of depthwise separable convolution DP and computational effort F DP as follows:
[0067] P DP =K 2 ×M+M×N (3)
[0068] F DP =K 2 ×H out ×W out ×M+H out ×W out ×M×N (4)
[0069] Therefore, the ratio of the number of parameters P and the number of computations F of standard convolution and depth-separable convolution are as follows:
[0070]
[0071] Where H out and W out They represent the height and width of the output feature map respectively, M represents the number of channels of the input feature map, K represents the convolution kernel size, and N represents the number of convolution kernels, which is also the number of channels of the output feature map. It can be seen from equations (5) and (6) that the number of parameters and computational complexity of depthwise separable convolution are significantly reduced compared to standard convolution.
[0072] The lightweight multi-scale feature extraction module first uses 1×1 convolution to map the feature map after passing through the fully connected layer from a low-dimensional space to a high-dimensional space to extract more information; then uses two feature branches with convolution scales of 3×3 and 5×5, and finally splices them on the channel, thereby achieving the acquisition of multi-scale features while maintaining a low computational load.
[0073] The bottleneck feature discrimination module is as follows: first, a layer normalization operation is performed on the bottleneck feature, and then a partial convolution is performed to achieve a preliminary discrimination between valid and invalid pixels; then, a second feature discrimination is performed through a gating mechanism controlled by a depthwise separable convolution; while discriminating local features, it can effectively reduce the number of parameters and computational complexity, and prevent the network from generating gradient explosion problems by introducing a residual structure.
[0074] The implementation process of the MFD-LRASPP vessel segmentation network constructed by single training in step (3) is as follows:
[0075] The data-enhanced features are used as the input of the convolutional network for a convolution operation with kernel_size=1, and then the feature vector S1 is obtained through the Relu6 activation function and layer normalization;
[0076] Entering the first Stage block, S1 is input into the Inverted Residual Block. S1 first passes through a convolution layer with 16 convolution kernels, the convolution kernel size is 3×3, the stride is 1, and then through batch normalization and layer normalization, and then through depth-separable convolution to obtain S11. S11 then passes through a layer of adaptive average pooling, 1×1 convolution, Relu activation function, 1×1 convolution and HardSigmoid activation layer to obtain S12. S11 and S12 are multiplied by a 1×1 convolution to obtain S2. Finally, S1 and S2 are added to obtain the feature vector S3. At this time, the feature vector is converted from the input (460, 460, 3) to (240, 240, 16); S3 also enters the next Inverted Residual Block with a stride of 2 to obtain S4 (240, 240, 16), and S4 enters the Inverted Residual Block with a stride of 1 Block obtains S5 (120, 120, 24); the fourth layer of the network is a lightweight multi-scale feature extraction module LMFB, S5 passes through three branches with a convolution stride of 2, among which branch one adopts 1×1 convolution, batch normalization and Relu activation function to obtain S51, branch two adopts 1×1 convolution, batch normalization, Relu activation function and 3×3 convolution to obtain S52, branch three adopts 1×1 convolution, batch normalization, Relu activation function and 5×5 convolution to obtain S53, and then the feature vectors obtained by the three branches are spliced along the channel dimension to finally obtain S6 (120, 120, 24); similarly, S6 passes through two layers of LMFB with a convolution stride of 1 to obtain S7 (60, 60, 40) and S8 (60, 60, 40); S8 then passes through the Inverted Residual with a convolution stride of 2 and an activation function of Hard Sigmoid Block obtains S9 (60, 60, 40); S9 is subjected to Inverted Residual Block with a convolution stride of 1 and an activation function of Hard Sigmoid to obtain S10 (30, 30, 40); S10 is subjected to two layers of LMFB with a convolution stride of 1 and an activation function of Hard Sigmoid to obtain S11 (30, 30, 40) and S12 (30, 30, 40); S12 is subjected to Inverted Residual Block with a convolution stride of 1 and an activation function of Hard Sigmoid to obtain S13 (30, 30, 40); S13 is subjected to four layers of Inverted Residual Block with a convolution stride of 1, an activation function of Hard Sigmoid and Dilated rata = [1, 2, 2, 2] Block obtains S14(30, 30, 112), S15(30, 30, 112), S16(30, 30, 160) and S17(30, 30, 160);S17 finally obtains bottleneck feature S18 (30, 30, 160) through a 1×1 convolution. Bottleneck feature S18 enters the bottleneck feature discrimination module (FDB), which first performs a layer normalization operation on the bottleneck feature and then performs a partial convolution to initially distinguish between valid and invalid pixels. Then, a gating mechanism controlled by a depthwise separable convolution performs a second feature discrimination to obtain S181. S18 and S181 are added to obtain the final backbone feature S19. S19 is sent to the LR-ASPP head for image segmentation.
[0077] The verification and evaluation process in step (4) is as follows: the constructed Yangtze River vessel recognition network based on the MFD-LRASPP network is trained, and the network model is iterated multiple times. The training parameters of the model are adjusted: learning rate (lr), batch size (batch_size), number of iterations (epochs), and optimizer (optimizer) until the optimal model parameters are obtained. The best training model weight is used as the pre-trained model to verify the data images in the validation set, and the performance of the model is evaluated using four indicators: accuracy, precision, recall, and mean average prediction (mAP).
[0078] The calculation formulas for each indicator are as follows:
[0079]
[0080] Where TP (True Positives) is a true positive example, which is predicted as a positive example and is actually a positive example; FP (False Positives) is a false positive example, which is predicted as a positive example but is actually a negative example; FN (False Negatives) is a false negative example, which is predicted as a negative example but is actually a positive example; TN (True Negatives) is a true negative example, which is predicted as a negative example but is actually a negative example. C represents the total number of categories, and AP represents the total number of categories. i Represents the AP value of the i-th class.
[0081] The image prediction in step (5) is as follows: First, the optimized network model weights are loaded to ensure optimal prediction accuracy. Then, the image to be processed is input into the deep learning model, which analyzes the image based on its learned features. Finally, the model outputs a vessel segmentation result, which accurately identifies and segments the vessel area in the image.
Claims
1. The Yangtze River ship recognition method based on MFD-LRASPP network is characterized by: The following steps are involved: (1) Segment the pre-filmed Yangtze River ship videos and create a dataset, and then preprocess the obtained Yangtze River ship dataset; (2) Construct a vessel segmentation network based on the MFD-LRASPP model; (3) Training the constructed ship segmentation network based on the MFD-LRASPP model, and adjusting the model training parameters multiple times until the optimal network model weight is obtained; (4) Verify and evaluate the performance of the obtained optimal network model weights; (5) The optimal network model weights after verification and evaluation are loaded, and the real surface ship navigation images are predicted to finally obtain the ship segmentation results.
2. The Yangtze River surface vessel identification method based on the MFD-LRASPP network according to claim 1 is characterized in that: The data set in step (1) is divided into a training set and a validation set in a ratio of 8:2; the preprocessing includes normalization and scaling.
3. The Yangtze River surface vessel identification method based on the MFD-LRASPP network according to claim 1 is characterized in that: The implementation process of pre-shooting the video of the ships on the Yangtze River in step (1) is as follows: First, the video was captured using a drone placed over the Yangtze River's waterways. The video parameters were a resolution of 3840×2160 and a frame rate of 60, with each waterway divided into sections every 200 meters. Secondly, we captured images at 1-second intervals from the video, filtered out images with small content differences, and labeled the ships in the images using the labelme annotation tool. Finally, the obtained json annotation file is converted into a segmentation label mask image, and the dataset is divided into a training set and a validation set.
4. The Yangtze River surface vessel identification method based on the MFD-LRASPP network according to claim 1 is characterized in that: The ship segmentation network in step (2) includes an inverted residual module, a lightweight multi-scale feature extraction module, a bottleneck feature discrimination module and a segmentation head LRASPP Head.
5. The Yangtze River surface vessel identification method based on the MFD-LRASPP network according to claim 4 is characterized in that: The inverted residual module adopts depth-wise separable convolution, including channel-by-channel convolution and point-by-point convolution; The number of parameters P of standard convolution s and computational effort F s as follows: P s =K 2 ×M×N (1) F s =H out ×W out ×M×K 2 ×N (2) The number of parameters P of depthwise separable convolution DP and computational effort F DP as follows: P DP =K 2 ×M+M×N (3) F DP =K 2 ×H out ×W out ×M+H out ×W out ×M×N (4) Therefore, the ratio of the number of parameters P and the number of computations F between standard convolution and depthwise separable convolution are as follows: Where H out and W out They represent the height and width of the output feature map, M represents the number of channels of the input feature map, K represents the convolution kernel size, and N represents the number of convolution kernels.
6. The method for identifying ships on the Yangtze River based on the MFD-LRASPP network according to claim 4 is characterized in that: The lightweight multi-scale feature extraction module is as follows: first, a 1×1 convolution is used to map the feature map after the fully connected layer from a low-dimensional space to a high-dimensional space; then, two feature branches with convolution scales of 3×3 and 5×5 are used, and finally, they are spliced on the channel to obtain multi-scale features.
7. The method for identifying ships on the Yangtze River based on the MFD-LRASPP network according to claim 4 is characterized in that: The bottleneck feature discrimination module is as follows: first, a layer normalization operation is performed on the bottleneck feature, and then a partial convolution is performed; then, a second feature discrimination is performed through a gating mechanism controlled by a depth-wise separable convolution.
8. The method for identifying ships on the Yangtze River based on the MFD-LRASPP network according to claim 1 is characterized in that: The implementation process of single training of the ship segmentation network in step (3) is as follows: the features after data enhancement are used as the input of the convolutional network to perform a convolution operation with kernel_size=1, and then the feature vector S1 is obtained by performing a Relu6 activation function and layer normalization; First, in the backbone network, input data S1 is processed through a 3×3 convolution layer, batch normalization, and layer normalization to generate S11. S11 is then activated through adaptive average pooling, 1×1 convolution, ReLU, and Hard Sigmoid to generate S12. The dot product of the two is then processed through a 1×1 convolution to form S2, which is then added to S1 to obtain the feature vector S3. S3 is processed through multiple inverted residual modules to generate S4 and S5. In the lightweight multi-scale feature extraction module, S5 extracts features through three different convolution branches and concatenates them into S6, which is then processed through two layers of LMFB to obtain S7 and S8. Next, S9 and S10 are generated in sequence through the inverted residual module, and then S11 and S12 are obtained through two layers of LMFB. S12 passes through multiple inverted residual modules with different expansion rates to finally generate S14-S17. Then, in the bottleneck layer, S17 obtains the bottleneck feature S18 through 1×1 convolution, and is optimized in the bottleneck feature discrimination module to generate S181 and add it to S18 to form the final feature S19, which is sent to the LR-ASPP Head for image segmentation.
9. The method for identifying ships on the Yangtze River based on the MFD-LRASPP network according to claim 1, characterized in that: The performance of the optimal network model weight obtained in step (4) is verified and evaluated as follows: the Yangtze River surface vessel recognition network based on the constructed MFD-LRASPP network is trained, and the network model is iteratively trained multiple times, and the training parameters of the model are adjusted: learning rate, batch size, number of iterations and optimizer, until the optimal model parameters are obtained; The best training model weight is used as the pre-training model to verify the data images in the validation set, and the performance of the model is evaluated by four indicators: accuracy, precision, recall rate and mAP; The calculation formulas for each indicator are as follows: In the formula, TP represents true positive examples, FP represents false positive examples, FN represents false negative examples, TN represents true negative examples, C represents the total number of categories, AP represents the total number of categories, and i represents the AP value of the i-th class.
10. The method for identifying ships on the Yangtze River based on the MFD-LRASPP network according to claim 1 is characterized in that: The implementation process of predicting the navigation image of a surface vessel in step (5) is as follows: first, loading the optimized network model weights to ensure the best prediction accuracy; then, inputting the image to be processed into the deep learning model, which analyzes the image based on its learned features; Finally, the model outputs the ship segmentation result, which accurately identifies and segments the ship area in the image.