Eel fry fishing investigation method based on high-resolution remote sensing image data
By improving the YOLOv8 model and combining high-resolution remote sensing image data, the problem of difficult monitoring of eel seedling fishing activities in China has been solved, efficient identification of eel seedling fishermen has been achieved, and monitoring accuracy and management capabilities have been improved.
Patent Information
- Application Number
- CN202411091229.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-22
AI Technical Summary
The existing technology is difficult to effectively monitor and manage China's coastal eel seedling fishing activities, especially in the Yangtze River estuary, and there are stealing behaviors, which makes it difficult to grasp the intensity and scale of eel seedling fishing.
Using the eel seedling fishing survey method based on high-resolution remote sensing image data, the high-resolution remote sensing image eel seedling fisherman float sample data set is established by improving the algorithm's ability to identify small targets, and using C2f_Dual, SPPELAN and AFPN The modules are used to enhance the model performance and achieve efficient identification of eel seedling fishermen.
It significantly improves the identification accuracy and recall rate of eel seedling fishermen, improves the average accuracy of 4.5% and the accuracy of 9.9%, and can more effectively monitor eel seedling fishing activities and provide stronger management support.
Smart Images

Figure CN119007029A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of marine resource fishing investigation and monitoring, and particularly to an eel fry fishing investigation method based on high-resolution remote sensing image data. Background Art
[0002] Japanese eel is a catadromous fish with high nutritional value. Adult eels migrate from freshwater habitats to the ocean to reproduce. The migration of eel fry is a very complex and precise biological phenomenon involving multiple environmental factors such as ocean currents, temperature, salinity, etc. Eel fry can sense the earth's magnetic field and use this ability to assist their migration process. Although artificial breeding of eels has become quite mature, the full life cycle breeding from eggs to adult fish has not been successfully achieved so far. The source of fry for eel farming still mainly relies on the capture of wild eel fry. Therefore, the intensity of eel fry fishing is crucial to the sustainable development of eel resources, especially the development of wild populations. Although China's fishery management department has adopted a special licensing system for the capture of eel fry, trying to control the scale and intensity of eel fry fishing, due to the possibility of poaching, people still know little about the scale of eel fry fishing in the Yangtze River Estuary and along the coast of China.
[0003] In recent years, remote sensing technology has made significant progress. The resolution of satellite sensors has been continuously improved, and it is now possible to provide images with sub-meter or even higher resolution. For example, my country's Jilin-1 wide-band 01C satellite can provide images with a resolution of up to 0.5 meters. This makes it possible to use high-resolution remote sensing images to carry out dynamic monitoring of eel fishing. In view of the difficulty in monitoring eel fishing along the coast of China, especially in the Yangtze River estuary, surveys and monitoring of eel fishing based on high-resolution remote sensing image data will help managers understand the spatial distribution of eel fishing activities and identify possible illegal fishing activities. With the development of artificial intelligence technology, especially the breakthrough of deep learning in the field of image recognition, the application of object recognition based on deep learning on high spatial resolution remote sensing images has changed the previous remote sensing survey that mainly relied on spectral and pixel-based information extraction, and at the same time provided a new solution for identifying eel fishing and conducting monitoring. Summary of the invention
[0004] The present invention provides a method for investigating eel fry fishing based on high-resolution remote sensing image data. By improving the YOLOv8 model, a high-resolution remote sensing image eel fry fishing device float sample dataset is established and experiments are carried out, which can be used for evaluating the intensity of eel fry fishing.
[0005] The present invention is achieved through the following technical solutions:
[0006] An eel fry fishing survey method based on high-resolution remote sensing image data.
[0007] First, obtain high spatial resolution (0.5m, or better than 0.5m resolution) remote sensing images that have been preprocessed by geometric correction and cover the survey area; second, label the image data of high spatial resolution remote sensing images and establish a high-resolution remote sensing image eel fisher float sample data set; then, improve the target detection algorithm YOLOv8 to improve the algorithm's recognition ability for small remote sensing targets, where C2f_Dual is used to replace four C2f modules in the backbone, SPPELAN replaces SPPF, and AFPN is combined with the original detection head at the Neck end to obtain four detection heads, and a 160×160 small target detection head is added, which improves the model's detection accuracy for small targets; finally, use the improved model to identify eel fishers in high spatial resolution remote sensing images covering the survey area and obtain the final results. The specific steps of YOLOv8 model improvement include the following:
[0008] S1. Combining DualConv and C2f into a lightweight structure
[0009] Efficient convolution filters can effectively reduce the computational cost and parameters of neural networks, and there is no need to design a new convolutional network architecture from scratch. Depthwise separable convolution was introduced in MobileNetV1, which significantly reduced the parameters and computation by decomposing the standard convolution into depthwise convolution and point convolution. Then, group convolution was introduced to reduce the computational cost but limit the information exchange between channels. Subsequently, heterogeneous convolution proposed to use convolution kernels of different sizes at the same time to reduce the computational complexity. Finally, Dual Convolutional Kernels combined the advantages of group convolution and heterogeneous convolution to design a new convolution filter.
[0010] S2. Improve and add a small target layer on the Neck side
[0011] In order to overcome the limitations of single-scale features, a multi-scale feature pyramid was constructed to transfer high-level features to low-level features through a top-down path to achieve multi-scale feature fusion; PAFPN added a bottom-up path to enhance the ability to obtain details of high-level features; NASFPN optimized the feature connection structure by using neural architecture search; GraphFPN introduced graph neural networks to achieve direct interaction of non-adjacent scale features, but increased the computational complexity; the combination of FPT and CARAFE with self-attention mechanism and context feature aggregation further improved the performance and wide application of feature pyramids;
[0012] S3. Improve SPPF using SPPELAN module
[0013] The SPPELAN module is a key component in YOLOv9, which achieves more efficient feature extraction and fusion by combining spatial pyramid pooling (SPP) and enhanced path aggregation network (ELAN).
[0014] Preferably, M is the number of input channels (i.e., the depth of the input feature map), N is the number of convolution filters, which is also the number of output channels (i.e., the depth of the output feature map), G is the number of groups in the group convolution and dual convolution, and the symbol The result of the 3×3 and 1×1 convolution kernels is added.
[0015] As a preferred method, DualConv combines 3×3 and 1×1 convolution kernels and uses group convolution technology to efficiently arrange convolution filters, which greatly reduces the computational cost and number of parameters of deep neural networks. This structural design enables DualConv to significantly reduce computational complexity while maintaining or improving model accuracy.
[0016] Preferably, 3×3 convolution moves in the feature map channel dimension, while 1×1 convolution is performed on all input channels, which can be seen as a combination of 3×3 group convolutions and 1×1 point-by-point convolutions on the same input feature map. Since DualConv applies 1×1 convolutions on all channels of the input feature map, the original information is retained, helping subsequent convolutional layers to extract features more fully. This enables DualConv to greatly reduce the computational cost and number of parameters while maintaining efficient feature extraction and information fusion, demonstrating its great potential and broad application prospects in the design of lightweight deep neural networks.
[0017] As a preferred method, when improving the YOLOv8 network structure, the C2f_Dual module introduces DualConv to replace the second ordinary convolution in the Bottleneck module based on the C2f module, which reduces the computational complexity, reduces the number of parameters, enhances the feature extraction capability, improves the model robustness and improves the small target detection capability, providing many benefits for improving the model performance and optimizing the practical application.
[0018] As a preferred method, YOLOv8 improves the detection capability of targets of different sizes through multi-scale feature fusion. The three detection heads set in the basic setting extract information from feature maps of three scales accordingly and fuse these features together to capture more contextual information during the detection process.
[0019] As a preferred choice, according to the data distribution characteristics of the eel fry fishing float on the high-resolution remote sensing image (0.5m), most small targets are smaller than 32×32 pixels. The current network cannot meet the detection capabilities of small targets, and false detection and missed detection are prone to occur in small target detection. Asymptotic Feature Pyramid Network (AFPN), a novel multi-scale feature pyramid network, is introduced to solve the problem of multi-scale feature extraction in target detection tasks. Compared with the original detection network, a new 160×160 feature map is added to correspond to a smaller detection head, thereby improving the detection accuracy of small targets.
[0020] As a preference, AFPN avoids the loss or degradation of feature information by directly interacting features between non-adjacent levels. The network first fuses two adjacent low-level features and gradually incorporates high-level features into the fusion process, thereby avoiding large semantic gaps between non-adjacent levels. In order to mitigate the conflicts that may arise in the feature fusion process of multi-object information at each spatial position, AFPN further uses adaptive spatial fusion operations to alleviate these inconsistencies.
[0021] Preferably, the adaptive feature fusion (ASFF) module used in AFPN is used to fuse feature maps of different scales to improve detection accuracy. Taking two-layer feature fusion as an example, the weight feature map formula 1 is first calculated as follows:
[0022]
[0023] Where X 1 and X 2 is the input feature map, represents the concatenation operation of the feature map, Conv represents the convolution operation, and Softmax represents the softmax activation function. Next, the formula 2 for calculating the fused feature map is as follows:
[0024] Y = Conv(X 1 ·W 1 +X 2 ·W 2 ) (2)
[0025] Where W 1 and W 2 is the weighted feature map, and · represents element-by-element multiplication. Similarly, the fusion of three-layer and four-layer feature maps is consistent with that of two layers.
[0026] As a preferred method, the SPPELAN module extracts multi-scale information through spatial pyramid pooling, which effectively improves the detection ability of targets of different sizes. The introduction of ELAN enhances the transmission of path information and feature fusion, enabling the network to better capture target features in complex scenes. Combining these two methods, SPPELAN provides richer contextual information, thereby improving the accuracy and robustness of target detection. The optimized network structure reduces the computational complexity while maintaining efficient feature extraction capabilities.
[0027] Technical principle of the present invention: This application takes the paired and single floats of eel fry catchers as the research objects, uses high spatial resolution remote sensing images to identify and mark the floats of eel fry catchers, and establishes a sample data set of eel fry catcher floats of high-resolution remote sensing images; in view of the problem of poor small target detection, an improved remote sensing small target detection algorithm of YOLOv8 is proposed. The algorithm first replaces the original detection head with AFPN, which has the advantage of small target detection, and then introduces C2f_Dual convolutional network to replace C2f to reduce the number of parameters, and uses SPPELAN to improve SPPF to improve the network detection speed and ensure the detection accuracy. Experimental tests were carried out on the sample data set of eel fry catcher floats made with Jilin No. 1 high-resolution remote sensing data as the original data. Compared with the original YOLOv8, the algorithm proposed in this application has an increase of 4.5% in mAP, 9.9% in precision (P), and 4.7% in recall (R), which has a good detection effect. This application demonstrates the great potential of high spatial resolution remote sensing data in small target recognition and monitoring of human fishing activities. If future applications and research can expand the coverage of remote sensing data, the present invention will have greater application value. The scheme of the present invention will provide more powerful technical support for the monitoring and management of eel fry fishing activities.
[0028] The present invention inherits and improves the previous YOLO series models. The core features of the algorithm of the network structure of the present invention include major improvements in the structure of the backbone network and the detection head, which improves the detection ability of small objects and the accuracy of the overall model. In the backbone part, C2f_Dual is used to replace four C2f modules, SPPELAN replaces SPPF, and AFPN is combined with the original detection head at the Neck end to obtain four detection heads, and a new 160×160 small target detection head is added. These improvements make it have better recognition ability for small targets than the original YOLOv8 model.
[0029] Beneficial effects:
[0030] The present invention uses high spatial resolution remote sensing images to establish a data set, which well solves the problem that small targets of sea surface floats are easily affected by waves and difficult to identify, and the present invention further improves the ability to detect small targets using high-resolution remote sensing images by improving the YOLOv8 model. The present invention introduces an AFPN detection head to add a small target detection layer, uses the Dual Convolutional Kernels module to improve C2f to reduce the computational cost and number of parameters of the deep neural network, significantly improves the recognition ability of small targets, and replaces the SPPF module with the SPPELAN module in YOLOv9 to enhance the feature extraction ability and improve the model speed to improve the YOLOv8 model. The present invention significantly enhances the ability to identify the small target of the eel fry fishing device float in the surface environment of the Yangtze River Estuary from high spatial resolution remote sensing images. The average accuracy mAP@0.5 of the algorithm proposed by the present invention is 4.5% higher than that of the original YOLOv8 algorithm, reaching 94.8%. Compared with the original model, the improved YOLOv8 has better positioning ability and detection accuracy for floats, and can evaluate the intensity of eel fry fishing. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 Schematic diagram of the original YOLOv8 network structure.
[0032] Figure 2 The figure is a schematic diagram of an improved YOLOv8 network structure in one embodiment of the present invention.
[0033] Figure 3 The figure is a schematic diagram of a double convolution convolution filter design in one embodiment of the present invention.
[0034] Figure 4 FIG. 4 is a schematic diagram of the structure of a C2f_Dual module in an embodiment of the present invention.
[0035] Figure 5 Schematic diagram of asymptotic feature pyramid network architecture in one embodiment of the present invention.
[0036] Figure 6 It is a schematic diagram of the SPPELAN structure in one embodiment of the present invention.
[0037] Figure 7 This is an F1 score diagram in one embodiment of the present invention, with the improved YOLOv8 on the left and the YOLOv8 on the right.
[0038] Figure 8 This is a schematic diagram of the accuracy curve in one embodiment of the present invention, with the improved YOLOv8 on the left and the YOLOv8 on the right.
[0039] Fig. 9The present invention is a schematic diagram showing a comparison of the detection results of a double float and a single float of an eel fry catcher before and after improvement in an embodiment of the present invention.
[0040] Fig.10 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0041] The following is a detailed description of an embodiment of the present invention in conjunction with the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.
[0042] like Figure 1 As shown in the figure, YOLOv8 is an advanced target detection algorithm that inherits and improves the previous YOLO series models. Figure 1 The core features of the algorithm include major improvements in the structure of the backbone network and the detection head, which improves the detection ability of small objects and the accuracy of the overall model. Compared with YOLOv5 and YOLOv7, YOLOv8 introduces more complex network structure changes, such as the structural change from C3 to C2f. Such changes increase the gradient flow and help to better learn complex features. In addition, YOLOv8 introduces a decoupled head to make classification and position prediction more independent, which improves the overall detection performance. It also uses an Anchor-Free method, which simplifies the complexity of the model and may increase the detection speed compared to the previous Anchor-Based method. In terms of the loss function, YOLOv8 uses Distribution Focal Loss to improve the model's recognition ability for objects of various sizes. In terms of training strategies, YOLOv8 draws on the strategies of other latest algorithms such as YOLOX and YOLOv7, especially in data enhancement and positive and negative sample matching technology, adopts Task-Aligned Assigner and closes Mosaic enhancement, and further improves the generalization ability and accuracy of the model in various environments. According to the scaling factor of the network, five models of different scales are provided: N / S / M / L / X. These improvements enable YOLOv8 to show superior performance in the field of target detection, especially in real-time application scenarios.
[0043] An eel fry fishing survey method based on high-resolution remote sensing image data.
[0044] First, obtain remote sensing images with high spatial resolution (0.5m, or better than 0.5m resolution) covering the survey area after preprocessing such as geometric correction; second, label the remote sensing images with high spatial resolution and establish a high-resolution remote sensing image eel catcher float sample dataset; then, improve the target detection algorithm YOLOv8 to improve the algorithm's recognition ability for small remote sensing targets, and the network structure (such as Figure 2 As shown in the figure, C2f_Dual is used to replace four C2f modules in the backbone, SPPELAN replaces SPPF, AFPN is combined with the original detection head at the Neck end to obtain four detection heads, and a 160×160 small target detection head is added. The detection accuracy of the model for small targets is improved; finally, the improved model is used to identify eel fry traps on high spatial resolution remote sensing images covering the survey area and obtain the final results.
[0045] This embodiment uses my country's independent Jilin-1 wide-width 01C satellite imagery to obtain remote sensing images with a spatial resolution of 0.5m after preprocessing such as geometric correction and after fusion of panchromatic and multispectral images covering the survey area. The high spatial resolution of the images makes it possible to identify the floats of fishing devices that are easily affected by waves. ArcGIS Pro software is used to identify and mark the floats of eel traps in the preprocessed remote sensing images, and a high-resolution remote sensing image eel trap float sample dataset (ManMiao) is established.
[0046] The specific steps of improving the YOLOv8 model include the following:
[0047] S1. Combining DualConv and C2f into a lightweight structure
[0048] Efficient convolution filters can effectively reduce the computational cost and parameters of neural networks, and there is no need to design a new convolutional network architecture from scratch. Initially, standard convolution was mainly used for feature extraction and channel fusion. Although it has powerful performance, it is computationally intensive. Subsequently, depthwise separable convolution was introduced in MobileNetV1, which significantly reduced parameters and computational complexity by decomposing the standard convolution into depthwise convolution and point convolution. Next, group convolution was introduced to reduce computational costs but limit information exchange between channels. Subsequently, heterogeneous convolution proposed the simultaneous use of convolution kernels of different sizes to reduce computational complexity. Finally, Dual Convolutional Kernels combines the advantages of group convolution and heterogeneous convolution to design new convolution filters (such as Figure 3 shown).
[0049] Among them, M is the number of input channels (i.e., the depth of the input feature map), N is the number of convolution filters, which is also the number of output channels (i.e., the depth of the output feature map), G is the number of groups in group convolution and dual convolution, and the symbol ⊕ is the sum of the results of 3×3 and 1×1 convolution kernels. DualConv combines 3×3 and 1×1 convolution kernels and uses group convolution technology to efficiently arrange convolution filters, greatly reducing the computational cost and number of parameters of deep neural networks. This structural design enables DualConv to significantly reduce computational complexity while maintaining or improving model accuracy. Secondly, 3×3 convolution moves in the feature map channel dimension, while 1×1 convolution is performed on all input channels. It can be regarded as a combination of 3×3 group convolution and 1×1 point-by-point convolution on the same input feature map. Since DualConv applies 1×1 convolution on all channels of the input feature map, it effectively retains the original information and helps subsequent convolution layers to extract features more fully. These enable DualConv to greatly reduce the computational cost and number of parameters while maintaining efficient feature extraction and information fusion, demonstrating its great potential and broad application prospects in the design of lightweight deep neural networks.
[0050] like Figure 4 As shown, when improving the YOLOv8 network structure, the present application is that the C2f_Dual module introduces DualConv to replace the second ordinary convolution in the Bottleneck module based on the C2f module, thereby reducing computational complexity, reducing the number of parameters, enhancing feature extraction capabilities, improving model robustness, and improving small target detection capabilities, providing many benefits for improving model performance and optimizing practical applications.
[0051] S2. Improve and add a small target layer on the Neck side
[0052] Traditional computer vision methods usually extract single-scale features from images for analysis and processing. However, this method performs poorly for objects of different sizes or scenes of different scales. In order to overcome the limitations of single-scale features, researchers have constructed multi-scale feature pyramids. The development of feature pyramids began with the classic FPN, which transfers high-level features to low-level features through a top-down path to achieve multi-scale feature fusion; PAFPN adds a bottom-up path to enhance the ability to obtain details of high-level features; NASFPN uses neural architecture search to optimize feature connection structures; GraphFPN introduces graph neural networks to achieve direct interaction of non-adjacent scale features, but increases computational complexity; other methods such as FPT and CARAFE combine self-attention mechanisms and contextual feature aggregation to further improve the performance and wide application of feature pyramids.
[0053] YOLOv8 improves the detection ability of objects of different sizes through multi-scale feature fusion. The three detection heads set up in the basic setting extract information from feature maps of three scales accordingly and fuse these features together to capture more contextual information during the detection process.
[0054] However, according to the data distribution characteristics of eel fry fishing floats on high-resolution remote sensing images (0.5m), most small targets are smaller than 32×32 pixels. The current network cannot meet the detection capabilities of small targets, and it is easy to have false detection and missed detection in small target detection. In this paper, the Asymptotic Feature Pyramid Network (AFPN, such as Figure 5 A novel multi-scale feature pyramid network is shown in Figure 1. It is used to solve the problem of multi-scale feature extraction in target detection tasks. Compared with the original detection network, a new 160×160 feature map is added to correspond to a smaller detection head and improve the detection accuracy of small targets.
[0055] AFPN avoids the loss or degradation of feature information by directly interacting features between non-adjacent levels. The network first fuses two adjacent low-level features and gradually incorporates high-level features into the fusion process, thereby avoiding large semantic gaps between non-adjacent levels. In addition, in order to mitigate the conflicts that may arise in the feature fusion process of multi-object information at each spatial position, AFPN further utilizes adaptive spatial fusion operations to alleviate these inconsistencies.
[0056] The adaptive feature fusion (ASFF) module used in AFPN is used to fuse feature maps of different scales to improve detection accuracy. Taking two-layer feature fusion as an example, the weight feature map formula 1 is first calculated as follows:
[0057]
[0058] Where X 1 and X 2 is the input feature map, represents the concatenation operation of the feature map, Conv represents the convolution operation, and Softmax represents the softmax activation function. Next, the formula 2 for calculating the fused feature map is as follows:
[0059] Y = Conv(X 1 ·W 1 +X 2 ·W 2 ) (4)
[0060] Where W 1 and W 2is the weighted feature map, and · represents element-by-element multiplication. Similarly, the fusion of three-layer and four-layer feature maps is consistent with that of two layers.
[0061] S3. Improve SPPF using SPPELAN module
[0062] like Figure 6 As shown in the figure, the SPPELAN module is a key component in YOLOv9. By combining spatial pyramid pooling (SPP) and enhanced path aggregation network (ELAN), more efficient feature extraction and fusion are achieved. It extracts multi-scale information through spatial pyramid pooling, effectively improving the detection ability of objects of different sizes. The introduction of ELAN enhances the transmission of path information and feature fusion, enabling the network to better capture target features in complex scenes. Combining these two methods, SPPELAN provides richer contextual information, thereby improving the accuracy and robustness of target detection. The optimized network structure reduces the computational complexity while maintaining efficient feature extraction capabilities.
[0063] The embodiments of the present application conducted experiments to evaluate the accuracy of the improved model. The experimental results and analysis are as follows:
[0064] (1) Experimental configuration
[0065] The experimental environment is based on Python 3.10.13 and implemented using the PyTorch2.1.0 framework in Windows 11. The processor used is the 11th generation Intel(R) Core(TM) i5-11400H, and the NVIDIA GeForce RTX3060GPU is used for training and testing. During image preprocessing, the height and width of the image are set to 640×640, and the batchsize is 16. The stochastic gradient descent optimizer (SGD) is used, with an initial learning rate of 0.01, a final learning rate of 0.001, a momentum of 0.937, and a weight decay of 0.0005. The hyperparameters are set to epochs=300, batch=8, workers=0, and iou=0.5.
[0066] (2) Evaluation indicators
[0067] The embodiments of the present application use indicators such as Precision, Recall, and mAP to evaluate the performance of the model.
[0068] 1. Precision refers to the ratio of the number of correctly detected targets to the number of all detected targets. The higher the precision, the fewer false positives the model has. The calculation formula is as follows:
[0069] Precision=TP / (TP+FP), (5)
[0070] TP is the true positive, that is, the number of correctly detected targets; FP is the false positive, that is, the number of incorrectly detected targets.
[0071] 2. Recall
[0072] It indicates the ratio of the number of correctly detected targets to the number of actually existing targets. The higher the recall rate, the fewer cases of missed detection by the model. The calculation formula is as follows:
[0073] Recall=TP / (TP+FN), (6)
[0074] Among them, TP is the true positive, that is, the number of targets detected correctly; FN is the false negative, that is, the number of targets that failed to be detected.
[0075] 3. Mean average precision (mAP) is a commonly used indicator to measure the overall performance of the target detection algorithm. mAP is the average of the average precision (AP) of different categories. Average precision (AP): For a certain category, calculate the area of the precision and recall curves under different thresholds. Mean average precision (mAP): Take the average of the AP of all categories. mAP50-95 means the average of the average precision (AP) calculated under multiple thresholds with IoU thresholds ranging from 0.50 to 0.95 (with a step size of 0.05). This indicator provides a more comprehensive evaluation than a single IoU threshold (such as mAP50), and can more accurately reflect the detection performance of the model at different degrees of overlap. The calculation formula is as follows:
[0076]
[0077] Among them, C represents the number of categories, AP c represents the average precision of the cth class.
[0078] Through these indicators, the performance of the model in the target detection task can be comprehensively evaluated.
[0079] (3) Through the above experiments, we can conclude that the overall performance of the improved YOLOv8 model is:
[0080] The embodiment of the present application first trained the original YOLOv8 model and the improved YOLOv8 model on the eel fry catcher float sample data set ManMiao. The results of the two methods on the test set are shown in the following table. The results show that the improved YOLOv8 model is 4.7%, 9.9%, 4.5% and 4.7% higher than the original YOLOv8 in recall (R), precision (P), mAP, and mAP50-95, respectively. Through these comparisons, we can preliminarily conclude that the improved YOLOv8 performs better in detecting small remote sensing targets in this application, and is superior to the original YOLOv8 in precision and accuracy.
[0081] Table 1 Performance comparison between YOLOv8 and the improved YOLOv8 algorithm of the present invention
[0082]
[0083] The embodiments of the present application respectively train and test the original YOLOv8 model and the improved YOLOv8 model under the same data set and experimental conditions. Figure 7 As shown in the figure, the F1 score of the model before the improvement is 0.88 with some fluctuations, while the F1 score of the improved model is improved to 0.93 and is smoother. This shows that the improved model is more stable in performance and more effectively captures target features in the feature extraction stage, especially the detection accuracy of small targets is significantly improved.
[0084] For the YOLOv8 model, its PR curve reflects the detection performance of the model under different confidence thresholds. Since the YOLO series models are known for their fast and efficient characteristics, the PR curve of YOLOv8 can usually show a high precision rate at a high recall rate, which means that it can effectively detect a large number of targets while reducing false detections. Figure 8 It can be seen that under the same recall rate, the PR curve of the YOLOv8 model improved by this application shows a higher precision rate in most threshold ranges, the curve is closer to the upper right corner as a whole, and the area under the curve (AUC) is larger, which indicates that the model has better comprehensive detection performance under different thresholds. This further shows that our improvement of the YOLOv8 model is effective and can improve the accuracy of remote sensing small target recognition.
[0085] In order to check the impact of different network structure changes on the model, an ablation experiment was conducted, and the results are shown in the following table. By replacing the detection head with AFPN with progressive feature fusion, the effect is significant, and the mAP result is improved from 90.3% to 94.2%; at the same time, replacing C2f with C2f_Dual reduces the computational cost and number of parameters of the convolution operation, and also improves the detection speed; on this basis, applying SPPELAN in the YOLOv9 algorithm to replace SPPF not only improves the accuracy, but also improves the accuracy and robustness of target detection, and further improves mAP to 94.8%. These results all show that the improved YOLOv8 model of this application effectively improves the performance of the original YOLOv8.
[0086] Table 2 Ablation experiment results
[0087]
[0088] like Fig. 9 As shown in the figure, it is a comparison of the detection results of the double float and single float of the eel fry catcher before and after the improvement of the YOLOv8 model. Fig. 9 a, 9b, and 9c are unimproved detection images. Fig. 9 d, 9e, and 9f are the improved detection images. It can be seen that there were missed detections in the red circles before the improvement. The improved model can solve the problem of missed detections of small targets well, and the overall detection results have also been improved, which also proves the effectiveness of the model improvement.
[0089] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A method for eel fry fishing survey based on high-resolution remote sensing image data, characterized in that: First, obtain high spatial resolution (0.5m, or better than 0.5m resolution) remote sensing images that have been preprocessed by geometric correction and cover the survey area; second, label the image data of high spatial resolution remote sensing images and establish a high-resolution remote sensing image eel fisher float sample data set; then, improve the target detection algorithm YOLOv8 to improve the algorithm's recognition ability for small remote sensing targets, use C2f_Dual to replace four C2f modules in the backbone, SPPELAN to replace SPPF, and use AFPN to combine with the original detection head at the Neck end to obtain four detection heads, and add a 160×160 small target detection head, so that the model's detection accuracy for small targets is improved; finally, use the improved model to identify eel fishers in high spatial resolution remote sensing images covering the survey area and obtain the final results. The specific steps of YOLOv8 model improvement include the following: S1. Combining DualConv and C2f into a lightweight structure Efficient convolution filters can effectively reduce the computational cost and parameters of neural networks, and there is no need to design a new convolutional network architecture from scratch. Depthwise separable convolution was introduced in MobileNetV1, which significantly reduced the parameters and computation by decomposing the standard convolution into depthwise convolution and point convolution. Then, group convolution was introduced to reduce the computational cost but limit the information exchange between channels. Subsequently, heterogeneous convolution proposed to use convolution kernels of different sizes at the same time to reduce the computational complexity. Finally, Dual Convolutional Kernels combined the advantages of group convolution and heterogeneous convolution to design a new convolution filter. S2. Improve and add a small target layer on the Neck side In order to overcome the limitations of single-scale features, a multi-scale feature pyramid was constructed to transfer high-level features to low-level features through a top-down path to achieve multi-scale feature fusion; PAFPN added a bottom-up path to enhance the ability to obtain details of high-level features; NASFPN optimized the feature connection structure by using neural architecture search; GraphFPN introduced graph neural networks to achieve direct interaction of non-adjacent scale features, but increased the computational complexity; the combination of FPT and CARAFE with self-attention mechanism and context feature aggregation further improved the performance and wide application of feature pyramids; S3. Improve SPPF using SPPELAN module The SPPELAN module is a key component in YOLOv9, which achieves more efficient feature extraction and fusion by combining spatial pyramid pooling (SPP) and enhanced path aggregation network (ELAN).
2. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: M is the number of input channels (i.e., the depth of the input feature map), N is the number of convolution filters, which is also the number of output channels (i.e., the depth of the output feature map), G is the number of groups in the group convolution and dual convolution, and the symbol ⊕ is the sum of the results of the 3×3 and 1×1 convolution kernels.
3. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: DualConv combines 3×3 and 1×1 convolution kernels and uses group convolution technology to efficiently arrange convolution filters, greatly reducing the computational cost and number of parameters of deep neural networks. This structural design enables DualConv to significantly reduce computational complexity while maintaining or improving model accuracy.
4. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: The 3×3 convolution moves in the feature map channel dimension, while the 1×1 convolution is performed on all input channels, which can be seen as a combination of 3×3 group convolutions and 1×1 point-by-point convolutions on the same input feature map. Since DualConv applies 1×1 convolutions on all channels of the input feature map, the original information is retained, helping subsequent convolutional layers to extract features more fully. This enables DualConv to greatly reduce the computational cost and number of parameters while maintaining efficient feature extraction and information fusion, demonstrating its great potential and broad application prospects in the design of lightweight deep neural networks.
5. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: When improving the YOLOv8 network structure, the C2f_Dual module introduces DualConv to replace the second ordinary convolution in the Bottleneck module based on the C2f module, reducing computational complexity, reducing the number of parameters, enhancing feature extraction capabilities, improving model robustness, and improving small target detection capabilities, providing many benefits for improving model performance and optimizing practical applications.
6. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: YOLOv8 improves the detection ability of objects of different sizes through multi-scale feature fusion. The three detection heads set up in the basic setting extract information from feature maps of three scales accordingly and fuse these features together to capture more contextual information during the detection process.
7. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: According to the data distribution characteristics of eel fry fishing floats on high-resolution remote sensing images (0.5m), most small targets are smaller than 32×32 pixels. The current network cannot meet the detection capabilities of small targets, and false detection and missed detection are prone to occur in small target detection. Asymptotic Feature Pyramid Network (AFPN), a novel multi-scale feature pyramid network, is introduced to solve the problem of multi-scale feature extraction in target detection tasks. Compared with the original detection network, a new 160×160 feature map is added to correspond to a smaller detection head, thereby improving the detection accuracy of small targets.
8. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: AFPN avoids the loss or degradation of feature information by directly interacting features between non-adjacent layers. The network first fuses two adjacent low-level features and gradually incorporates high-level features into the fusion process, thereby avoiding large semantic gaps between non-adjacent layers. In order to mitigate the conflicts that may arise in the feature fusion process of multi-object information at each spatial position, AFPN further uses adaptive spatial fusion operations to alleviate these inconsistencies.
9. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: The adaptive feature fusion (ASFF) module used in AFPN is used to fuse feature maps of different scales to improve detection accuracy. Taking two-layer feature fusion as an example, the weight feature map formula 1 is first calculated as follows: Where X1 and X2 are input feature maps, represents the concatenation operation of the feature map, Conv represents the convolution operation, and Softmax represents the softmax activation function. Next, the formula 2 for calculating the fused feature map is as follows: Y=Conv(X1·W1+X2·W2) (9) Where W1 and W2 are weighted feature maps, and · represents element-by-element multiplication. Similarly, the fusion of three-layer and four-layer feature maps is consistent with that of two layers.
10. The method for eel catching investigation based on high-resolution remote sensing image data according to claim 1, characterized in that: The SPPELAN module extracts multi-scale information through spatial pyramid pooling, effectively improving the detection ability of targets of different sizes. The introduction of ELAN enhances the transmission of path information and feature fusion, enabling the network to better capture target features in complex scenes. Combining these two methods, SPPELAN provides richer contextual information, thereby improving the accuracy and robustness of target detection. The optimized network structure reduces the computational complexity while maintaining efficient feature extraction capabilities.
Citation Information
Cited By
Fish behavior analysis device and method based on underwater image processing
CN120032423A
An apparatus and method for fish behavior analysis based on underwater image processing
CN120032423B
