Sea surface enteromorpha detection method and system based on deep learning

Abstract

The invention belongs to the technical field of environment monitoring, and particularly relates to a sea surface enteromorpha detection method and system based on deep learning, and the method comprises the steps: firstly introducing a ConvNeXtv2 module into a backbone network of YOLOv8 to improve the feature extraction capability of a model, and facilitating the extraction of the edge features of enteromorpha in a complex environment; secondly, an ECA attention mechanism is introduced into the Neck network, and the perception ability of the model to different scales is enhanced, so that the detection precision is improved; and finally, replacing a traditional loss function with an EIoU loss function, optimizing the positioning capability of the model prediction frame, and further improving the identification effect of the enteromorpha.

Description

Technical Field

[0001] The present application belongs to the field of environmental monitoring technology, and specifically relates to a method and system for detecting sea surface enteromorpha based on deep learning. Background Art

[0002] In recent years, outbreaks of Enteromorpha have severely impacted the socioeconomic and marine ecosystems along the Yellow Sea coast. Traditional methods for monitoring Enteromorpha rely primarily on manual inspections, satellite and aerial remote sensing, and fixed-point monitoring. These methods suffer from poor real-time performance, insufficient accuracy, and limited monitoring range, making them difficult to effectively monitor and warn of Enteromorpha disasters.

[0003] With the development of artificial intelligence and computer vision technologies, deep learning-based object detection methods are gradually being applied in environmental monitoring. In particular, the YOLO series of object detection algorithms, with their excellent real-time performance and high efficiency, are widely used in a variety of object detection tasks.

[0004] However, Enteromorpha communities lack distinct internal features, their edges shift over time, and they can aggregate, disperse, settle, and disappear. Furthermore, they are significantly affected by external factors such as light and currents. This results in low accuracy and a high rate of missed detections when dealing with the complex background of the ocean and the unique morphology of Enteromorpha, as exemplified by the traditional YOLOv8 model. Summary of the Invention

[0005] Based on the above problems, the present application provides a method and system for detecting sea surface enteromorpha based on deep learning. This method replaces the Bottleneck module in C2f with the ConvNeXtv2 module to form a new C2f-ConvNeXtv2 module. Compared with the Bottleneck module, the ConvNeXtv2 module uses a large-kernel deep separable convolution, which reduces the amount of calculation while expanding the model's perception ability and can capture long-distance dependencies across objects. In addition, the GRN layer introduced by ConvNeXtv2 realizes cross-channel feature calibration, realizes competition and global perception between channels, and solves the problem of redundancy of traditional convolution channels. Its technical content is as follows:

[0006] A method for detecting enteromorpha on the sea surface based on deep learning, comprising the following steps:

[0007] S1. Collect images of Enteromorpha;

[0008] S2. Image preprocessing;

[0009] S3. Annotate the Enteromorpha dataset;

[0010] S4. Input the labeled dataset into the improved YOLOv8 model, which includes the backbone network, the feature fusion network, and the detection head.

[0011] The backbone network Backbone extracts features from the input Enteromorpha image and fuses feature map information from different layers and scales through the feature fusion network Neck. The detection head Head completes the image recognition task on feature maps of different scales through decoupling design and dynamic label allocation strategy, and finally outputs the location information of Enteromorpha in the image.

[0012] Preferably, the image preprocessing steps in step S2 are as follows:

[0013] S2.1. Perform brightness processing on the acquired image;

[0014] S2.2. Perform contrast processing on the acquired image;

[0015] S2.3. Add Gaussian noise to the collected image.

[0016] Preferably, in step S3, the Enteromorpha dataset is annotated using annotation software, and the specific operations are as follows:

[0017] S3.1. Use annotation software to draw a rectangular box around each Enteromorpha leaf in the image. If there are multiple Enteromorpha leaves in the image, mark each leaf with a separate rectangular box.

[0018] S3.2. The Enteromorpha image dataset and the corresponding label dataset are divided into training set, validation set, and test set in a ratio of 8:1:1. The training set is used to train the model, the validation set is used to tune the model during training, and the test set is used to evaluate the generalization ability of the final model.

[0019] Preferably, in step S4, the YOLOv8 network is improved to obtain CEE-YOLOv8, which specifically includes the following contents:

[0020] (4.1) Replace the Bottleneck module in C2f with the ConvNeXtv2 module, and replace the C2f modules in layers 2, 4, 6, and 8 of the Backbone network with C2f-ConvNeXtv2 modules;

[0021] (4.2) The ECA attention mechanism is introduced after the 15th, 18th, and 21st layers of the Neck network;

[0022] (4.3) Replace the original CIoU with the EIoU loss function.

[0023] Preferably, the backbone network Backbone in (4.1) performs feature extraction on the input Enteromorpha image, and the specific steps are as follows:

[0024] S.4.1.1. The ConvNeXtv2 network first extracts features from the edge of the Enteromorpha using a 7x7 depthwise separable convolution. Depthwise separable convolution splits the traditional convolution operation into two stages: depthwise convolution and pointwise convolution. Depthwise convolution first performs independent convolution on each channel and then combines the output features using pointwise convolution. S4.1.2. A LN layer is used to normalize the feature maps of each channel, and then a 1x1 convolution kernel is used to enhance the features of the Enteromorpha image edges.

[0025] S4.1.3. Use the GELU activation function instead of the ReLU function;

[0026] S4.1.4. ConvNeXtv2 introduces a GRN layer, which achieves inter-channel competition and global perception with zero parameter cost through purely mathematical constraints. The initial features input to ConvNeXtv2 are fused with the extracted features through a residual structure to obtain a fused feature map. Deeper feature extraction is performed in the ConvNeXtv2 module.

[0027] S4.1.5. The fused feature map output by the last ConvNeXtv2 module is concatenated with the output of the secondary branch split by the initial split module along the channel dimension. The concatenated feature map is input to the SPPF module for maximum pooling to fuse the features of Enteromorpha at multiple scales.

[0028] Preferably, the feature fusion network Neck in (4.2) performs feature fusion in the following specific steps:

[0029] S4.2.1. The feature map after pooling in the SPPF module enters the feature fusion network Neck. After operations such as Upsample, Concat, and C2f, the feature map F enters the ECA attention mechanism.

[0030] S4.2.2. In the ECA attention module, first the input feature map F of each channel is Perform global average pooling to compress the global spatial information of each channel into a value called the channel descriptor z c , the formula is as follows:

[0031]

[0032] Where H represents the height of the input feature map F, W represents the width of the input feature map F, C represents the number of input channels, i and j represent the values ​​of the i-th row and j-th column of the input feature map F; the obtained C channel descriptors will form a C-dimensional column vector S4.2.3. Use one-dimensional convolution to process the compressed channel column vector z to obtain a set of processed column vectors s, which is formulated as follows:

[0033] s = Conv1D(z,k);

[0034] Conv1D stands for 1D convolution. The convolution kernel k of 1D convolution will be adaptively adjusted according to the number of channels. The formula is as follows:

[0035] ψ(·) represents an adaptive function, where C represents the number of input channels, and γ and b represent hyperparameters. After the one-dimensional convolution, the column vector s is mapped through the Sigmoid function to obtain a set of attention weights w, as shown in the following formula:

[0036] w=σ(s)=σ(Conv1D(z,k));

[0037]

[0038] The attention weight w is between 0 and 1. The closer the value is to 1, the more important the current task of the channel is.

[0039] S4.2.4. Attention weights Multiply the input feature map F channel by channel to get the output feature map F' c :

[0040] w c represents the attention weight of the c-th channel; F c represents the input feature map of the cth channel; F c ' represents the output feature map of the c-th channel.

[0041] Preferably, (4.3) EIoU loss function is calculated as follows:

[0042] L EIoU =L IoU +L distance +L aspect ;

[0043] L IoU = 1-IoU;

[0044]

[0045] Among them, IoU measures the degree of overlap between the real box and the predicted box, A represents the area of ​​the predicted box, and B represents the area of ​​the real box; L distance Represents the center point distance loss, which is used to measure the Euclidean distance between the center point of the predicted box and the real box; L aspect represents the aspect ratio loss, which is used to optimize the difference in width and height between the predicted box and the real box; ρ(·) represents the Euclidean distance between the center of the predicted box and the real box, l represents the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box, w and h are the width and height of the predicted box; w gt and h gt is the width and height of the real box; l w and l h is the width and height of the minimum bounding rectangle, d represents the center coordinate of the prediction box, d gt Indicates the center coordinates of the real box.

[0046] A deep learning-based sea surface enteromorpha detection system, comprising an image acquisition unit, an image processing unit, and an output unit;

[0047] Image acquisition unit: Use an unmanned boat equipped with an image acquisition gimbal to shoot videos of Enteromorpha on the sea surface and extract images of Enteromorpha on the sea surface frame by frame;

[0048] Image processing unit: annotates the Enteromorpha dataset and trains the improved YOLOv8 model;

[0049] Output unit: Use the trained model to monitor the Enteromorpha on the sea surface and provide visual output.

[0050] Compared with the prior art, this application has the following beneficial effects:

[0051] 1. This method replaces the Bottleneck module in C2f with the ConvNeXtv2 module, forming the new C2f-ConvNeXtv2 module. Compared to the Bottleneck module, the ConvNeXtv2 module uses depthwise separable convolution with a large kernel. This reduces computational complexity while expanding the model's perception and enabling it to capture long-range dependencies across objects. Furthermore, the GRN layer introduced in ConvNeXtv2 enables cross-channel feature calibration, enabling inter-channel competition and global perception, and resolving the channel redundancy issue in traditional convolutional models.

[0052] 2. The ECA attention mechanism is introduced into the Neck network. The ECA module obtains a C-dimensional vector through operations such as global average pooling and one-dimensional convolution. Channel attention weights are then obtained through sigmoid function mapping. This effectively enhances the signal strength of important channels and suppresses secondary channels, allowing the model to focus more on areas related to Enteromorpha information, thereby improving the model's detection accuracy for long-distance Enteromorpha.

[0053] 3. At the same time, the introduction of the EIoU loss function improves the positioning ability of the prediction box, enhances the recognition ability of the model, and makes the Enteromorpha recognition algorithm more stable and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of CEE-YOLOv8;

[0055] Figure 2 Schematic diagram of C2f-ConcNextv2;

[0056] Figure 3 This is a schematic diagram of the C2f module;

[0057] Figure 4 This is a schematic diagram of ConvNeXtv2;

[0058] Figure 5 Schematic diagram of the ECA attention mechanism module;

[0059] Figure 6 This is the traditional YOLOv8 recognition result diagram;

[0060] Figure 7 This is the identification result diagram for this application. DETAILED DESCRIPTION

[0061] The technical solution of the present application is described in detail below through specific embodiments and drawings. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application, and the specific technical features can be combined with each other.

[0062] A method for detecting enteromorpha on the sea surface based on deep learning, comprising the following steps:

[0063] S1. Collect images of Enteromorpha;

[0064] S1.1. Collect images of Enteromorpha using an independently developed unmanned vessel.

[0065] The unmanned boat uses an efficient battery management system and can work continuously for 8-10 days. It is equipped with an image acquisition gimbal integrated with 6 cameras, which can achieve 360° acquisition without blind spots.

[0066] S1.2. Extract the collected Enteromorpha video frame by frame to obtain an Enteromorpha image;

[0067] First, since the resolution of the collected Enteromorpha images is 2560×1440, in order to save subsequent training time, the photos were cropped to preserve the Enteromorpha itself and some background information.

[0068] S2. Image preprocessing;

[0069] S2.1. Perform brightness processing on the collected images. Brightness processing helps improve the model's adaptability to lighting changes in different time periods and solves the brightness concentration problem caused by the concentrated collection time period.

[0070] S2.2. Perform contrast processing on the collected images. Improving the contrast can magnify the difference between the Enteromorpha and the background, enhance the visibility of details, and help the model better identify the Enteromorpha.

[0071] S2.3. Add Gaussian noise to the collected images. Adding noise can enhance the robustness of the model, prevent overfitting, and allow the model to adapt to changes in noise in the real world.

[0072] S3. Annotate the Enteromorpha dataset;

[0073] S3.1. Use annotation software to draw a rectangular box around each Enteromorpha leaf in the image. The box should tightly enclose the edges of the Enteromorpha leaf and avoid any excess background. If there are multiple Enteromorpha leaves in the image, label each leaf with its own rectangular box.

[0074] After the annotation is completed, a corresponding txt document will be generated. There are 5 parameters in the txt document: category, normalized center point coordinates (x, y), and normalized width and height of the target box.

[0075] S3.2. The Enteromorpha image dataset and the corresponding label dataset are divided into training set, validation set, and test set in a ratio of 8:1:1. The training set is used to train the model, the validation set is used to tune the model during training, and the test set is used to evaluate the generalization ability of the final model.

[0076] S4. Input the labeled dataset into the improved YOLOv8 model, which includes the backbone network, the feature fusion network, and the detection head.

[0077] The backbone network Backbone extracts features from the input Enteromorpha image and fuses feature map information from different layers and scales through the feature fusion network Neck. The detection head Head completes the image recognition task on feature maps of different scales through decoupling design and dynamic label allocation strategy, and finally outputs the location information of Enteromorpha in the image.

[0078] In step S4, the YOLOv8 network is improved to obtain CEE-YOLOv8. The specific method is as follows:

[0079] (4.1) Replace the Bottleneck module in C2f with the ConvNeXtv2 module, and replace the C2f modules in layers 2, 4, 6, and 8 of the Backbone network with C2f-ConvNeXtv2 modules;

[0080] The backbone network extracts features from the input Enteromorpha image. The specific steps are as follows:

[0081] In the C2f-ConvNeXtv2 module, the input feature map X1 is first passed through the Conv module to extract feature information, resulting in the feature map X2.

[0082] The S4.1.2.split module evenly splits the feature map X2 into two branches along the channel dimension. The main branch enters the added ConvNeXtv2 network for deep feature extraction. The secondary branch does not undergo any operation and retains the original information of the input feature map. Finally, it is spliced ​​with the feature map processed by ConvNeXtv2 through the residual structure;

[0083] In ConvNeXtv2, the feature map X2 first passes through a large 7×7 convolution kernel to obtain an intermediate feature map X3; the intermediate feature map X3 enters the LN layer,

[0084] The S4.1.4.LN layer normalizes the feature map of each channel. The normalized feature map undergoes a two-dimensional convolution operation to obtain the convolution feature map Y. The convolution feature map Y is then input into the GRN layer via the GELU activation function.

[0085] S4.1.5. In the GRN layer, L2 norm of the convolution feature map of each channel Perform aggregation processing to obtain a set of aggregate values ​​G(Y)=gy={||Y1||,||Y2||,...,||Y C ||}∈R C ;

[0086] Among them, G(Y) i =||Y i || is a scalar, which represents the aggregate value of the feature map of the i-th channel; H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels;

[0087] S4.1.6. Apply the standard segmentation normalization function N(·) to the above aggregated values ​​to obtain a set of normalized values;

[0088] S4.1.7. Use the calculated normalized score to multiply the convolution feature map Y channel by channel, Y i =Y i *N(G(Y)i )∈R H×W Get the reference feature map X4;

[0089] S4.1.8. The reference feature map X4 undergoes a two-dimensional convolution operation to further extract feature image information. Finally, the initial features input to ConvNeXtv2 are fused with the extracted features through a residual structure to obtain a fused feature map X5.

[0090] S4.1.9. Send the fused feature map to the next ConvNeXtv2 module for deeper feature extraction, and finally output a fused feature map X6 that is fused with feature map X5.

[0091] S4.1.10. Feature map X6 is concatenated with the output of the secondary branch from the initial split module along the channel dimension. A 1×1 convolution is then performed to adjust the channels so that the number of output channels equals the number of input channels.

[0092] S4.1.11. Finally, the spliced ​​feature map X7 is output and input into the SPPF module for maximum pooling to fuse the features of Enteromorpha at multiple scales.

[0093] (4.2) The ECA attention mechanism is introduced after the 15th, 18th, and 21st layers of the Neck network;

[0094] The specific steps of feature fusion network Neck are as follows:

[0095] S4.2.1. The feature map after pooling in the SPPF module enters the feature fusion network Neck. After operations such as Upsample, Concat, and C2f, the feature map F enters the ECA attention mechanism.

[0096] S4.2.2. In the ECA attention module, first the input feature map F of each channel is Perform global average pooling to compress the global spatial information of each channel into a value called the channel descriptor z c , the formula is as follows:

[0097]

[0098] Where H represents the height of the input feature map F, W represents the width of the input feature map F, C represents the number of input channels, i and j represent the values ​​of the i-th row and j-th column of the input feature map F; the obtained C channel descriptors will form a C-dimensional column vector

[0099] S4.2.3. Use one-dimensional convolution to process the compressed channel column vector z to obtain a set of processed column vectors s, which is formulated as follows:

[0100] s = Conv1D(z,k);

[0101] Conv1D stands for 1D convolution. The convolution kernel k of 1D convolution will be adaptively adjusted according to the number of channels. The formula is as follows:

[0102] ψ(·) represents the adaptive function, where C represents the number of input channels, γ and b represent hyperparameters (usually set to γ ​​= 2, b = 1), |·| odd It means taking the odd number closest to the result;

[0103] The column vector s after 1D convolution is mapped through the Sigmoid function to obtain a set of attention weights w. The formula is as follows:

[0104] w=σ(s)=σ(Conv1D(z,k));

[0105]

[0106] The weight value is between 0 and 1. The closer the value is to 1, the more important the current task of the channel is.

[0107] S4.2.4. Attention weights Multiply the input feature map F channel by channel to get the output feature map F':

[0108] w c represents the attention weight of the c-th channel; F c represents the input feature map of the cth channel; F c ' represents the output feature map of the c-th channel.

[0109] (4.3) Replace the original CIoU with the EIoU loss function.

[0110] The EIoU loss function is calculated as follows:

[0111] L EIoU =L IoU +L distance +L aspect ;

[0112] L IoU = 1-IoU;

[0113]

[0114] Among them, IoU measures the degree of overlap between the real box and the predicted box, A represents the area of ​​the predicted box, and B represents the area of ​​the real box; L distance Represents the center point distance loss, which is used to measure the Euclidean distance between the center point of the predicted box and the real box; L aspect represents the aspect ratio loss, which is used to optimize the difference in width and height between the predicted box and the real box; ρ(·) represents the Euclidean distance between the center of the predicted box and the real box, l represents the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box, w and h are the width and height of the predicted box; w gt and h gt is the width and height of the real box; l w and l h is the width and height of the minimum bounding rectangle, d represents the center coordinate of the prediction box, d gt Indicates the center coordinates of the real box.

[0115] A deep learning-based sea surface enteromorpha detection system, comprising an image acquisition unit, an image processing unit, and an output unit;

[0116] Image acquisition unit: Use an unmanned boat equipped with an image acquisition gimbal to shoot videos of Enteromorpha on the sea surface and extract images of Enteromorpha on the sea surface frame by frame;

[0117] Image processing unit: annotates the Enteromorpha dataset and trains the improved YOLOv8 model;

[0118] Output unit: Use the trained model to monitor the Enteromorpha on the sea surface and provide visual output.

[0119] Enteromorpha distribution has irregular shapes and sizes. The EIoU loss function improves the accuracy of bounding box matching to Enteromorpha by optimizing the coverage distance, aspect ratio, and center point distance of the bounding box, avoiding the gradient vanishing problem of traditional IoU loss on non-overlapping bounding boxes and reducing the harmful gradient impact of low-quality anchor boxes during training.

[0120] Use the training set to train the CEE-YOLOv8 model, then use the validation set to tune the model and save the best training model.

[0121] To verify the feasibility of the proposed solution, a simulation experiment was conducted. The Enteromorpha recognition task was performed on a laptop running the Windows 10 operating system. The specific experimental computer configuration is shown in Table 1.

[0122] Table 1 Training environment configuration of Enteromorpha experiment

[0123] Configuration Specific Version CPU Intel(R)core(TM)i7-8700CPU@3.2GHZ GPU NVIDIA GeForce RTX 2080Ti Operation Windows 10 (64-bit) Editing language Python 3.8 Deep Learning framework Pytorch 1.12.1

[0124] During training, the batch size is set to 16, the learning rate is 0.01, the top epoch is 100, and the input image size is 640×640.

[0125] Step 6: Use the test set to test the trained CEE-YOLOv8 model, calculate the performance indicators of the model on the test set, and complete the detection accuracy evaluation of the model.

[0126] Specifically, the present invention uses four evaluation indicators: precision, recall, mean average precision (mAP50), and computational complexity to evaluate the detection accuracy and lightweight performance of the model. The calculation formula is as follows:

[0127]

[0128] Where T P Indicates the positive sample predicted to be positive, F P Denotes negative samples predicted as positive, F N Indicates a positive sample predicted as negative. N is the total number of target categories. There is only one category of Enteromorpha, so N=1, AP i is the average precision value of the i-th category. The computational cost refers to the number of floating-point operations required for the model to perform inference. The smaller the computational cost, the faster the model inference speed.

[0129] To demonstrate the advancement of our technical solution, we compared the training results of the CEE-YOLOv8 model with those of the original YOLOv8 model. Table 2 shows the performance comparison between the improved CEE-YOLOv8 model and the original model.

[0130] Table 2 Performance comparison of the improved model before and after

[0131]

[0132] As shown in Table 1, the improved model has an improvement of 2.6% in Precision, 1.2% in Recall, and 3.4% in mAP50-95. At the same time, the computational complexity is reduced by 0.7G. Furthermore, a qualitative comparison of the recognition results of the two models can provide an intuitive understanding of the detection performance of the two models. Figure 7 As shown in the figure, the original YOLOv8 model has a certain degree of missed detection in the recognition of Enteromorpha and has poor recognition performance at long distances. The improved CEE-YOLOv8 has greatly improved this situation and can accurately identify Enteromorpha communities at a distance.

[0133] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for detecting enteromorpha on the sea surface based on deep learning, characterized in that: The following steps are involved: S1. Collect images of Enteromorpha; S2. Image preprocessing; S3. Annotate the Enteromorpha dataset; S4. Input the labeled dataset into the improved YOLOv8 model, which includes the backbone network, the feature fusion network, and the detection head. The backbone network Backbone extracts features from the input Enteromorpha image and fuses feature map information from different layers and scales through the feature fusion network Neck. The detection head Head completes the image recognition task on feature maps of different scales through decoupling design and dynamic label allocation strategy, and finally outputs the location information of Enteromorpha in the image.

2. The method for detecting enteromorpha on the sea surface based on deep learning according to claim 1, characterized in that: The image preprocessing steps in step S2 are as follows: S2.

1. Perform brightness processing on the acquired image; S2.

2. Perform contrast processing on the acquired image; S2.

3. Add Gaussian noise to the collected image.

3. The method for detecting enteromorpha on the sea surface based on deep learning according to claim 1, characterized in that: In step S3, annotation software is used to annotate the Enteromorpha dataset. The specific operations are as follows: S3.

1. Use annotation software to draw a rectangular box around each Enteromorpha leaf in the image. If there are multiple Enteromorpha leaves in the image, mark each leaf with a separate rectangular box. S3.

2. The Enteromorpha image dataset and the corresponding label dataset are divided into training set, validation set, and test set in a ratio of 8:1:

1. The training set is used to train the model, the validation set is used to tune the model during training, and the test set is used to evaluate the generalization ability of the final model.

4. The method for detecting enteromorpha on the sea surface based on deep learning according to claim 1, characterized in that: In step S4, the YOLOv8 network is improved to obtain CEE-YOLOv8, which specifically includes the following contents: (4.1) Replace the Bottleneck module in C2f with the ConvNeXtv2 module, and replace the C2f modules in layers 2, 4, 6, and 8 of the Backbone network with C2f-ConvNeXtv2 modules; (4.2) The ECA attention mechanism is introduced after the 15th, 18th, and 21st layers of the Neck network; (4.3) Replace the original CIoU with the EIoU loss function.

5. The method for detecting enteromorpha on the sea surface based on deep learning according to claim 4, characterized in that: The backbone network in (4.1) extracts features from the input Enteromorpha image. The specific steps are as follows: S.4.1.

1. The ConvNeXtv2 network first extracts features from the edge of the Enteromorpha using a 7x7 depthwise separable convolution. Depthwise separable convolution splits the traditional convolution operation into two stages: depthwise convolution and pointwise convolution. Depthwise convolution first performs independent convolution on each channel and then combines the output features using pointwise convolution. S4.1.

2. Use the LN layer to normalize the feature map of each channel, and then use a 1×1 convolution kernel to enhance the edge features of the Enteromorpha image; S4.1.

3. Use the GELU activation function instead of the ReLU function; S4.1.

4. ConvNeXtv2 introduces a GRN layer, which achieves inter-channel competition and global perception with zero parameter cost through purely mathematical constraints. The initial features input to ConvNeXtv2 are fused with the extracted features through a residual structure to obtain a fused feature map. Deeper feature extraction is performed in the ConvNeXtv2 module. S4.1.

5. The fused feature map output by the last ConvNeXtv2 module is concatenated with the output of the secondary branch split by the initial split module along the channel dimension. The concatenated feature map is input to the SPPF module for maximum pooling to fuse the features of Enteromorpha at multiple scales.

6. The method for detecting enteromorpha on the sea surface based on deep learning according to claim 4, characterized in that: The specific steps of feature fusion in the feature fusion network Neck in (4.2) are as follows: S4.2.

1. The feature map after pooling in the SPPF module enters the feature fusion network Neck. After operations such as Upsample, Concat, and C2f, the feature map F enters the ECA attention mechanism. S4.2.

2. In the ECA attention module, first the input feature map F of each channel is Perform global average pooling to compress the global spatial information of each channel into a value called the channel descriptor z c , the formula is as follows: Where H represents the height of the input feature map F, W represents the width of the input feature map F, C represents the number of input channels, i and j represent the values ​​of the i-th row and j-th column of the input feature map F; the obtained C channel descriptors will form a C-dimensional column vector S4.2.

3. Use one-dimensional convolution to process the compressed channel column vector z to obtain a set of processed column vectors s, which is formulated as follows: s = Conv1D(z,k); Conv1D stands for 1D convolution. The convolution kernel k of 1D convolution will be adaptively adjusted according to the number of channels. The formula is as follows: ψ(·) represents an adaptive function, where C represents the number of input channels, and γ and b represent hyperparameters. After the one-dimensional convolution, the column vector s is mapped through the Sigmoid function to obtain a set of attention weights w, as shown in the following formula: w=σ(s)=σ(Conv1D(z,k)); S4.2.

4. Attention weights Multiply the input feature map F channel by channel to obtain the output feature map F′ c : w c represents the attention weight of the c-th channel; F c represents the input feature map of the cth channel; F c ' represents the output feature map of the c-th channel.

7. The method for detecting enteromorpha on the sea surface based on deep learning according to claim 4, characterized in that: (4.3) The EIoU loss function is calculated as follows: L EIoU =L IoU +L distance +L aspect ; L IoU =1-IoU; Among them, IoU measures the degree of overlap between the real box and the predicted box, A represents the area of ​​the predicted box, and B represents the area of ​​the real box; L distance Represents the center point distance loss, which is used to measure the Euclidean distance between the center point of the predicted box and the real box; L aspect represents the aspect ratio loss, which is used to optimize the difference in width and height between the predicted box and the real box; ρ(·) represents the Euclidean distance between the center of the predicted box and the real box, l represents the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box, w and h are the width and height of the predicted box; w gt and h gt is the width and height of the real box; l w and l h is the width and height of the minimum bounding rectangle, d represents the center coordinate of the prediction box, d gt Indicates the center coordinates of the real box.

8. A deep learning-based sea surface enteromorpha detection system, using the deep learning-based sea surface enteromorpha detection method according to any one of claims 1 to 7, characterized in that: It includes a picture acquisition unit, a picture processing unit and an output unit; Image acquisition unit: Use an unmanned boat equipped with an image acquisition gimbal to shoot videos of Enteromorpha on the sea surface and extract images of Enteromorpha on the sea surface frame by frame; Image processing unit: annotates the Enteromorpha dataset and trains the improved YOLOv8 model; Output unit: Use the trained model to monitor the Enteromorpha on the sea surface and provide visual output.

Images