A method for detecting maritime targets based on efficient channel-priority attention

Abstract

This invention relates to a method for detecting maritime targets based on efficient channel-priority attention, comprising: acquiring a maritime image to be detected and inputting it into a pre-constructed and trained maritime target detection model to obtain maritime target detection results. The maritime target detection model includes a backbone network, a neck network, and a multi-scale adaptive spatial feature fusion detection head connected in sequence. The backbone network is an improved RepViT backbone network, incorporating a compression excitation module with residuals and multi-scale depth-separable efficient channel-priority attention. The neck network is a YOLOv8 neck network with multiple scale feature outputs. The multi-scale adaptive spatial feature fusion detection head predicts multiple pre-selected boxes based on the multiple scale features output by the neck network, and determines the final target box and confidence score from the multiple pre-selected boxes. Compared with existing technologies, this invention can effectively improve the detection performance of maritime targets, reduce the search time of UAVs, and can be used in contexts such as maritime rescue.

Description

Technical Field

[0001] This invention relates to the field of maritime target detection technology, and in particular to a maritime target detection method based on efficient channel-priority attention. Background Technology

[0002] Marine activities involving the exploitation of marine resources inevitably face various emergencies, such as ship accidents and maritime disasters. After a disaster, timely search and rescue operations to ensure the safety of victims are crucial, necessitating a stronger maritime safety guarantee. Unlike land-based rescues, timeliness and efficiency are paramount at sea. Traditional maritime methods rely on manual, grid-like searches, which are inefficient over large areas, and stranded individuals tend to drift and disperse with ocean currents, making it difficult to pinpoint their exact location. Limited by human resources and response time, traditional methods have limitations in locating maritime incidents, impacting the timeliness of rescue operations. The rapid development of drone technology in recent years has provided powerful support for rescue efforts. During search and rescue operations, miniaturized and rapid drones can be equipped with various sensors to collect data on the changing environment, helping rescuers locate victims more quickly and significantly improving the efficiency of search and rescue operations.

[0003] Current drone-based maritime missions heavily rely on operators' visual identification of people in the water. However, drone cameras, under wide-area search conditions, produce small images of people in the water, resulting in blurry images. Both the human eye and traditional target detection algorithms have limitations, making accurate identification of people in the water impossible. Currently, deep learning-based target detection algorithms are developing rapidly, with common algorithms including SSD, Fast R-CNN, and the YOLO series. While these algorithms perform well in target detection, their effectiveness in detecting small targets is less than ideal. In complex images captured by drones, small targets constitute only a very small proportion and are easily confused with environmental information. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as the heavy reliance of operators on visual identification of people in the water in maritime missions involving drones, and the unsatisfactory performance of common target detection algorithms in detecting small targets. This invention provides a maritime target detection method based on efficient channel-priority attention.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for detecting maritime targets based on efficient channel-priority attention includes the following steps:

[0007] The system acquires a marine image to be detected and inputs it into a pre-built and trained marine target detection model to obtain marine target detection results. The marine target detection model includes a backbone network, a neck network, and a multi-scale adaptive spatial feature fusion detection head connected in sequence.

[0008] The backbone network is an improved RepViT backbone network, which introduces a compressed excitation module with residuals and a multi-scale depth-separable high-efficiency channel-priority attention.

[0009] The neck network is a YOLOv8 neck network with outputs of features at multiple different scales;

[0010] The multi-scale adaptive spatial feature fusion detection head predicts multiple pre-selected boxes based on the various scale features output by the neck network, and determines the final target box and confidence score from the multiple pre-selected boxes.

[0011] Furthermore, the training process of the maritime target detection model includes:

[0012] The images of people who fell into the water were collected and labeled by drones, and the images in the aerial photography dataset were preprocessed.

[0013] Furthermore, the preprocessing includes performing Mosaic data enhancement and Mixup data enhancement. The Mosaic data enhancement includes: randomly reading four images from the input image, performing rotation, scaling and color gamut transformation operations, then randomly arranging and stitching them together, and selecting a portion of the region as the newly generated image, while cropping the images outside the region.

[0014] The Mixup data augmentation involves randomly reading two images from the input image and stacking the two images and their corresponding labels together to obtain a mixed image and its corresponding labels.

[0015] Furthermore, the improved RepViT backbone network includes a Stem module, a Stage module, an SPPF module, and an ECPA module connected in sequence. There are multiple Stage modules, and a downsampling module is provided between each adjacent Stage module. The output of each Stage module is connected to the neck network.

[0016] Furthermore, the Stage module includes a RepViTSE-ResBlock module with a residual SE compression activation block. The RepViTSE-ResBlock module incorporates an SE-ResNet block to retain more feature information. The SE-ResNet block uses average pooling to compress the input feature map into 1×1×C, and then processes the compressed features through two fully connected layers, a ReLU activation function, and a Sigmoid activation function. The processed feature map is weighted according to the feature weights and added to the original input to obtain the output.

[0017] Furthermore, the processing procedure of the ECPA module includes: processing the input feature map through channel attention and multiplying it with the input feature map to obtain feature F'; taking feature F' as input, processing it through spatial attention and multiplying it with feature F' to obtain feature F'.

[0018] Furthermore, the channel attention performs cross-channel information interaction through one-dimensional convolution. The size of the convolution kernel of the one-dimensional convolution adaptively changes through a function k, the expression of which is:

[0019]

[0020] In the formula, γ = 2, b = 1, and c is the dimension of the input feature map;

[0021] The spatial attention is used to dynamically allocate attention weights in the channel and spatial dimensions, and the corresponding calculation expression is:

[0022]

[0023] In the formula, SA CPCA (F) represents the result after spatial attention processing on the input feature F, and DwConv(F) represents the depthwise convolution of the input feature F, where i∈{0,1,2,3} represents the i-th branch. i Conv indicates that the corresponding branch is selected. 1×1 This represents a 1×1 convolution.

[0024] Furthermore, the number of Stage modules is four, namely Stage1, Stage2, Stage3, and Stage4, arranged sequentially. The YOLOv8 neck network includes six stages. The first, second, and third stages each include an upsampling module, a Concat module, and a C2f module arranged sequentially. The first stage upsamples the output of the ECPA module and combines it with the output of the Stage3 module; the second stage upsamples the output of the first stage and combines it with the output of the Stage2 module; the third stage upsamples the output of the second stage and combines it with the output of the Stage1 module.

[0025] The fourth, fifth, and sixth stages each include a Conv convolution module, a Concat module, and a C2f module arranged in sequence. The fourth stage processes the output of the third stage using the Conv convolution module and combines it with the output of the second stage; the fifth stage processes the output of the fourth stage using the Conv convolution module and combines it with the output of the first stage; and the sixth stage processes the output of the fifth stage using the Conv convolution module and combines it with the output of the ECPA module.

[0026] The outputs of the C2f module in stages three, four, five, and six serve as four different scale features of the YOLOv8 neck network output.

[0027] Furthermore, the multi-scale adaptive spatial feature fusion detection head acquires multiple features of different scales from the output of the neck network, mixes them based on adaptive weights, and obtains feature information of multiple different scales for target detection.

[0028] Furthermore, the multi-scale adaptive spatial feature fusion detection head uses a slice-assisted super inference algorithm to predict multiple pre-selected boxes. The slice-assisted super inference algorithm divides the image to be detected into multiple sub-images and re-normalizes the sub-images into standard-sized images for target detection. The detection results are then processed by removing redundant detection boxes through non-maximum values ​​and converted into the coordinates of the original image to obtain the target detection results.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] (1) This invention is applicable to the scenario of detecting maritime rescue targets by maritime unmanned aerial vehicle (UAV) systems. The maritime rescue target detection algorithm provided by this invention effectively solves the problems of poor detection effect of small targets and weak anti-interference ability from the perspective of UAVs.

[0031] (2) The improved RepViT backbone network proposed in this invention uses the ECPA module to process input features through channel attention and multi-scale depth-separable convolutional spatial attention, thus completing cross-channel and cross-space attention of feature information; the Stage module incorporates SE-ResNet blocks that can retain more feature information, which can effectively enhance the feature information of small targets, reduce the false negative rate, and avoid the loss of necessary feature information; through the repeated downsampling module and the Stage module, smaller features can be passed down to and to the neck network; the overall backbone network can effectively enhance the feature information of small targets, reduce the false negative rate, improve the success rate of maritime rescue missions, and ensure the safety of personnel.

[0032] (3) The improved YOLOv8 neck network of this invention has four different scale feature outputs, which can fuse feature maps of different dimensions and output four different scale feature information to the prediction layer; the four-scale adaptive spatial feature fusion detection head can fuse four different scale feature information for target detection, and as a whole can pay more attention to the feature information of small targets and improve the model's ability to detect small targets.

[0033] (4) The four-scale adaptive spatial feature fusion detection head of the present invention uses a slice-assisted super inference algorithm to predict multiple pre-selected boxes, which can effectively improve the model's detection effect on small targets and improve the model's final detection performance.

[0034] (5) The Mosaic and Mixup hybrid data augmentation algorithm used in this invention enhances the diversity of data and improves the anti-interference ability and generalization ability of the detection model. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a method for detecting maritime targets based on efficient channel-priority attention provided in an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of an improved RepViT backbone network provided in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the structure of an ECPA module provided in an embodiment of the present invention;

[0038] Figure 4 This is a network structure diagram of a maritime target detection method based on efficient channel-priority attention provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0040] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0041] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0042] Example 1

[0043] This embodiment provides a method for detecting maritime targets based on efficient channel-priority attention, the process of which includes the following steps:

[0044] The system acquires images of the sea to be detected and inputs them into a pre-built and trained sea target detection model to obtain the sea target detection results. The sea target detection model includes a backbone network, a neck network, and a multi-scale adaptive spatial feature fusion detection head connected in sequence.

[0045] The backbone network is an improved RepViT backbone network, which introduces a compressed excitation module with residuals and a multi-scale depth-separable efficient channel-priority attention.

[0046] The neck network is a YOLOv8 neck network with outputs of features at multiple different scales;

[0047] The multi-scale adaptive spatial feature fusion detection head predicts multiple pre-selected boxes based on the various scale features output by the neck network, and determines the final target box and confidence score from the multiple pre-selected boxes.

[0048] like Figure 1 and Figure 4 As shown, the training process of the above method includes the following steps:

[0049] S1: Collect and annotate marine images using drones to obtain a dataset of aerial photographs of people who have fallen into the water;

[0050] In this implementation, the images collected are aerial images of people falling into the water taken from the perspective of a drone. The dataset is divided into a training set and a validation set in an 8:2 ratio.

[0051] S2: Preprocess the images in the aerial photography dataset of people who fell into the water to perform data augmentation;

[0052] In this embodiment, the preprocessing of the training set includes Mosaic data augmentation and Mixup data augmentation. Mosaic data augmentation includes: randomly reading four images from the input image, performing rotation, scaling and color gamut transformation operations, then randomly arranging and stitching them together, and selecting a portion of the region as the newly generated image, while cropping the images outside the region.

[0053] Mixup data augmentation involves randomly reading two images from the input image and stacking the two images and their corresponding labels together to obtain a mixed image and its corresponding labels;

[0054] Specifically, the training set is augmented with Mosaic and Mixup data at a ratio of 9:1. Mosaic data augmentation randomly reads four images from the current batch of input images, performs rotation, scaling, and color gamut transformation operations, randomly arranges and stitches these four images, and selects a 640×640 area as the newly generated image. Images outside the area are cropped.

[0055] Mixup data augmentation randomly reads two images from the current batch of input images and stacks the two images together with their ground truth bounding boxes according to the following formula:

[0056] λ = Beta(α,β)

[0057] mixed_batch = λ × batch image1 +(1-λ)×batch image2

[0058] Where λ follows a Beta distribution, image1 and image2 and their corresponding label values ​​are linearly transformed with λ as the coefficient to obtain the mixed image mixed_batch and its corresponding label, and then output.

[0059] S3: The improved RepViT backbone network, which incorporates a residual compression and excitation network (rSE) ​​and a multi-scale deep separable efficient channel-priority attention (ECPA), is used to extract feature information from the dataset images.

[0060] like Figure 2 As shown, the improved RepViT backbone network includes a Stem module, a Stage module, an SPPF module, and an ECPA module connected in sequence. There are multiple Stage modules, and a downsampling module is provided between each adjacent Stage module. The output of each Stage module is connected to the neck network.

[0061] Optionally, in this embodiment, there are four Stage modules, namely Stage1, Stage2, Stage3 and Stage4, which are distributed sequentially. The improved RepViT backbone network can therefore be divided into six stages, and the first stage, the second stage, the third stage and the sixth stage are selected as the feature outputs for the next level neck network.

[0062] The Stage module includes the RepViTSE-ResBlock module with residual SE compression activation blocks. The RepViTSE-ResBlock module incorporates an SE-ResNet block to retain more feature information. The SE-ResNet block uses average pooling to compress the input feature map into a 1×1×C shape. Then, the compressed features are processed through two fully connected layers, a ReLU activation function, and a Sigmoid activation function. The processed feature map is weighted according to the feature weights and added to the original input to obtain the output.

[0063] In addition, in this embodiment, the Stem module consists of two 3×3 convolutional and activation functions;

[0064] The downsampling module first uses the RepViTBlock module, then performs a 3×3 depthwise separable convolution with a stride of 2 to extract features, and finally completes downsampling after passing through a 1×1 unfolded convolution and an FFN module. The Spatial Pyramid Pooling–Fast (SPPF) module extracts features at different scales through a CBS convolution module and three 5×5 max pooling layers, and then achieves multi-scale feature fusion through channel merging.

[0065] The ECPA module is a multi-scale, depth-separable, high-efficiency channel-priority attention module based on the CABM attention framework, such as... Figure 3 As shown, the processing procedure includes: converting the input feature map F∈R... C*H*W Through channel attention M c Process and multiply with the input feature map to obtain the features Using feature F' as input, and applying spatial attention M... s Process and multiply with feature F' to obtain feature Complete cross-channel and cross-spatial focus on feature information;

[0066] Efficient Channel Attention (ECA) uses one-dimensional convolutions to facilitate information exchange across channels. The size of the convolution kernel adapts adaptively using a function k, allowing layers with a larger number of channels to perform more cross-channel interactions. The expression for the function k is:

[0067]

[0068] In the formula, γ = 2, b = 1, and c is the dimension of the input feature map;

[0069] Multi-scale depthwise separable convolutional spatial attention (CPCA) can dynamically allocate attention weights in both channel and spatial dimensions. It utilizes depthwise convolution to capture spatial relationships between features, ensuring that inter-channel relationships are preserved while reducing computational complexity. Its multi-scale structure enhances the ability of convolutional operations to capture spatial relationships. The computational description of multi-scale depthwise separable convolutional spatial attention is as follows:

[0070]

[0071] In the formula, SA CPCA (F) represents the result after spatial attention processing on the input feature F, and DwConv(F) represents the depthwise convolution of the input feature F, where i∈{0,1,2,3} represents the i-th branch. i Conv indicates that the corresponding branch is selected. 1×1 This represents a 1×1 convolution.

[0072] The specific processing steps of the above step S3 to improve the RepViT backbone network can be summarized as follows:

[0073] The improved RepViT backbone network consists of six stages. In the first stage, the input image is processed sequentially through the Stem module and the Stage module, and the width and height of the output feature map are reduced to one-quarter of the original.

[0074] In the second to fourth stages, the downsampling module and the Stage module are repeated to pass smaller-sized features down and up to the neck network.

[0075] In the fifth stage, the feature map is processed using the SPPF module. The feature map is first convolved, gradient eliminated by BN, and activated by the SiLU activation function in the SPPF module through the CBS module. Then it is passed through three consecutive 5×5 max pooling layers. The result of each pooling layer is concatenated with the feature map before pooling using a Concat tensor. The result is then processed by a CBS module to obtain the output.

[0076] In the sixth stage, the feature maps from the previous stage are processed using ECPA attention, and the results are output to the neck network.

[0077] S4: The feature information extracted by the backbone network is fed into the improved YOLOv8 neck network with multiple feature outputs at different scales, the feature maps of different dimensions are fused, and the feature information of multiple scales is output to the prediction layer.

[0078] In this embodiment, the YOLOv8 neck network is a network with four different scale feature outputs, including an upsample module, a concat module, a C2f module, and a convolutional Conv module. The overall framework is divided into six stages:

[0079] The first, second, and third stages have the same structure, each consisting of an upsampling module, a Concat module, and a C2f module arranged sequentially. The first stage upsamples the output of the ECPA module and combines it with the output of the Stage3 module; the second stage upsamples the output of the first stage and combines it with the output of the Stage2 module; and the third stage upsamples the output of the second stage and combines it with the output of the Stage1 module.

[0080] The fourth, fifth, and sixth stages each include a Conv convolution module, a Concat module, and a C2f module arranged in sequence. The fourth stage processes the output of the third stage using the Conv convolution module and combines it with the output of the second stage; the fifth stage processes the output of the fourth stage using the Conv convolution module and combines it with the output of the first stage; and the sixth stage processes the output of the fifth stage using the Conv convolution module and combines it with the output of the ECPA module.

[0081] The outputs of the C2f module in stages three, four, five, and six serve as four different scale features of the YOLOv8 neck network output.

[0082] The above six stages are equivalent to:

[0083] The first to third stages have the same structure. Each stage consists of an upsampling module, a Concat module, and a C2f module arranged in sequence. The Concat module in the first stage receives the output of the upsampling module and the output of the Stage3 module in the feature extraction backbone network. Correspondingly, the Concat module in the second stage receives the output of the Stage2 module in the feature extraction backbone network, and the Concat module in the third stage receives the output of the Stage1 module in the feature extraction backbone network.

[0084] The structure of stages four through six is ​​the same. Each stage consists of a Conv convolutional module, a Concat module, and a C2f module arranged in sequence. The Concat module in stage four receives the output of the Conv convolutional module as well as the output of the C2f module in stage two. Correspondingly, the Concat module in stage five receives the output of the C2f module in stage one, and the Concat module in stage six receives the output of the ECPA module in the feature extraction backbone network.

[0085] S5: Input the feature information of the four different scales output in step S4 into the Four-scale Adaptively Spatial Feature Fusion (FASFF) head to predict multiple pre-selected boxes, and determine the final target box and confidence score from the multiple pre-selected boxes.

[0086] Preferably, the multi-scale adaptive spatial feature fusion detection head acquires features of various scales from the output of the neck network, and mixes them based on adaptive weights to obtain feature information of various scales for target detection. The fusion formula is as follows:

[0087]

[0088] in, It is the four-layer feature information extracted from the neck network. This represents the feature information at the current layer scale, where i and j are the pixel position parameters, l is the current layer number, 2→l represents the path from the second layer to the current layer, and α, β, γ, and δ are weight parameters. The formulas for calculating the weight parameters are as follows:

[0089]

[0090] Among them, the parameters α, β, γ, and δ are concatenated by the tensor of the Concat module and then activated by the softmax function so that their range is within [0,1] and their sum is 1;

[0091] Preferably, the multi-scale adaptive spatial feature fusion detection head uses a slice-assisted super inference algorithm to predict multiple pre-selected boxes. The slice-assisted super inference algorithm divides the image to be detected into multiple sub-images and re-normalizes the sub-images into standard-sized images for target detection. The detection results are then processed by removing redundant detection boxes through non-maximum values ​​and converted into the coordinates of the original image to obtain the target detection results.

[0092] Essentially, the slice-assisted super inference algorithm is a method that cuts the image to be detected into overlapping small blocks for separate detection. The image to be detected is divided into small sub-images, which are then re-normalized into standard-sized images and fed into the model for detection or inference. The detection results are then processed by non-maximum suppression to remove redundant detection boxes and converted back into the coordinates of the original image. Finally, the detection results and inference results are merged and output.

[0093] The algorithm of this invention and the YOLOv8 algorithm were used simultaneously for target detection in marine images to compare their detection performance. The comparison results are shown in Table 1.

[0094] Table 1 Comparison of detection performance between the algorithm of this invention and the YOLOv8 algorithm

[0095]

[0096] As shown in Table 1, the algorithm of this invention is based on the YOLOv8 algorithm framework, but it has improved in all four sets of indicators compared with the YOLOv8 algorithm. The precision has been improved by 7.9%, the recall by 3.3%, the mAP@.5 by 7.1%, and the mAP@.5:.95 by 10.8%, indicating that the improved model algorithm has better detection performance than the original framework algorithm.

[0097] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for detecting maritime targets based on efficient channel-priority attention, characterized in that, Includes the following steps: The system acquires a marine image to be detected and inputs it into a pre-built and trained marine target detection model to obtain marine target detection results. The marine target detection model includes a backbone network, a neck network, and a multi-scale adaptive spatial feature fusion detection head connected in sequence. The backbone network is an improved RepViT backbone network, which introduces a compressed excitation module with residuals and a multi-scale depth-separable high-efficiency channel-priority attention. The neck network is a YOLOv8 neck network with outputs of features at multiple different scales; The multi-scale adaptive spatial feature fusion detection head predicts multiple pre-selected boxes based on the various scale features output by the neck network, and determines the final target box and confidence score from the multiple pre-selected boxes; The improved RepViT backbone network includes an ECPA module. The ECPA module's processing steps include: processing the input feature map through channel attention and multiplying it with the input feature map to obtain the feature map. , will feature As input, it is processed through spatial attention and combined with features Multiply to obtain the characteristics ; The channel attention mechanism facilitates cross-channel information exchange through one-dimensional convolution. The size of the convolution kernel adapts adaptively using a function k, the expression of which is: In the formula, γ=2, b=1, and c is the dimension of the input feature map; The spatial attention is used to dynamically allocate attention weights in the channel and spatial dimensions, and the corresponding calculation expression is: In the formula, For input features The result after spatial attention processing To perform a depthwise convolution on the input feature F, where i∈{0, 1, 2, 3} represents the i-th branch, This indicates that the corresponding branch has been selected. This represents a 1×1 convolution.

2. The maritime target detection method based on efficient channel-priority attention according to claim 1, characterized in that, The training process of the maritime target detection model includes: The images of people who fell into the water were collected and labeled by drones, and the images in the aerial photography dataset were preprocessed.

3. The maritime target detection method based on efficient channel-priority attention according to claim 2, characterized in that, The preprocessing includes Mosaic data enhancement and Mixup data enhancement. The Mosaic data enhancement includes: randomly reading four images from the input image, performing rotation, scaling and color gamut transformation operations, then randomly arranging and stitching them together, and selecting a portion of the image as the newly generated image, while cropping the images outside the specified area. The Mixup data augmentation involves randomly reading two images from the input image and stacking the two images and their corresponding labels together to obtain a mixed image and its corresponding labels.

4. The maritime target detection method based on efficient channel-priority attention according to claim 1, characterized in that, The improved RepViT backbone network also includes a Stem module, a Stage module, and an SPPF module connected in sequence. The ECPA module is connected after the SPPF module. There are multiple Stage modules, and a downsampling module is provided between each adjacent Stage module. The output of each Stage module is connected to the neck network.

5. A method for detecting maritime targets based on efficient channel-priority attention according to claim 4, characterized in that, The Stage module includes a RepViTSE-ResBlock module with a residual SE compression activation block. The RepViTSE-ResBlock module incorporates an SE-ResNet block to retain more feature information. The SE-ResNet block uses average pooling to compress the input feature map into 1×1×C, and then processes the compressed features through two fully connected layers, a ReLU activation function, and a Sigmoid activation function. The processed feature map is weighted according to the feature weights and added to the original input to obtain the output.

6. The maritime target detection method based on efficient channel-priority attention according to claim 4, characterized in that, The system comprises four Stage modules: Stage1, Stage2, Stage3, and Stage4, arranged sequentially. The YOLOv8 neck network consists of six stages. The first, second, and third stages each include an upsampling module, a Concat module, and a C2f module arranged sequentially. The first stage upsamples the output of the ECPA module and combines it with the output of Stage3. The second stage upsamples the output of the first stage and combines it with the output of Stage2. The third stage upsamples the output of the second stage and combines it with the output of Stage1. The fourth, fifth, and sixth stages each include a Conv convolution module, a Concat module, and a C2f module arranged in sequence. The fourth stage processes the output of the third stage using the Conv convolution module and combines it with the output of the second stage; the fifth stage processes the output of the fourth stage using the Conv convolution module and combines it with the output of the first stage; and the sixth stage processes the output of the fifth stage using the Conv convolution module and combines it with the output of the ECPA module. The outputs of the C2f module in stages three, four, five, and six serve as four different scale features of the YOLOv8 neck network output.

7. The maritime target detection method based on efficient channel-priority attention according to claim 1, characterized in that, The multi-scale adaptive spatial feature fusion detection head acquires multiple features of different scales from the output of the neck network, and mixes them based on adaptive weights to obtain feature information of multiple different scales for target detection.

8. A method for detecting maritime targets based on efficient channel-priority attention according to claim 7, characterized in that, The multi-scale adaptive spatial feature fusion detection head uses a slice-assisted super inference algorithm to predict multiple pre-selected boxes. The slice-assisted super inference algorithm divides the image to be detected into multiple sub-images and re-normalizes the sub-images into standard-sized images for target detection. The detection results are then processed by removing redundant detection boxes through non-maximum values ​​and converted into the coordinates of the original image to obtain the target detection results.

Images