An intelligent monitoring method and system for fish in rivers and lakes based on underwater robots

Abstract

The application discloses a kind of river and lake fish intelligent monitoring method and system based on underwater robot, by introducing ELA module on the basis of YOLOv8 framework, replace C2f in backbone network, neck network introduces LSKA module, ELA focuses on local area, refines target contour edge and detail, strengthens feature recognition degree;LSKA is covered with wide-area feature in dim scene Filter background noise, accurately locate the overall contour of the fixed target, avoid target missed detection.Combined with underwater robot, effectively solve the problem of limited accuracy of traditional monitoring method, general target detection algorithm transplantation application is prone to misdiagnosis and feature extraction accuracy is insufficient, by the technical design of adapting to underwater special environment, reduce water turbidity, uneven illumination and other factors interference, accurately identify fish target, guarantee the accuracy and effectiveness of monitoring data.

Description

Technical Field

[0001] This invention relates to the field of river and lake fish monitoring technology, and in particular to an intelligent monitoring method and system for river and lake fish based on an underwater robot. Background Technology

[0002] River and lake fish resources are an important component of aquatic ecosystems, and their population size, distribution characteristics, and activity patterns directly reflect the quality of the aquatic ecological environment. Monitoring and assessing river and lake fish resources is fundamental to protecting aquatic biodiversity and maintaining the ecological balance of rivers and lakes. It also serves as the basis for formulating fisheries management strategies and achieving the sustainable use of fishery resources. Furthermore, it provides basic data support for aquatic ecological research, and is of great significance in ecological environmental protection, fisheries industry development, and biodiversity research.

[0003] Currently, river and lake fish resource monitoring technologies are mainly divided into two categories: The first is traditional monitoring methods and conventional fixed underwater video monitoring technology. Traditional methods include manual fishing statistics and acoustic detection. Conventional fixed underwater video monitoring technology acquires underwater fish activity information by deploying imaging equipment at fixed points. Specifically, underwater camera groups with matching supplementary lighting equipment are installed in key river and lake sections, fish spawning grounds, migration channels, etc., to collect video data according to preset modes, which is then transmitted via wired or wireless communication to a shore-based control center or cloud server for storage. Subsequent manual playback, interpretation, and analysis are then used to extract fish information. The second category is the application of general target detection algorithms, including the transplantation of intelligent algorithms such as YOLOv8, which perform well in general scenarios, to the automated analysis of underwater video data to achieve fish target identification and information extraction. This type of technology is currently less widely used in the field of underwater fish monitoring.

[0004] Existing technologies have significant shortcomings: traditional and fixed video monitoring has low accuracy, interferes with the ecosystem, and lacks flexibility; manual interpretation is inefficient and prone to errors, and cannot support large-scale long-term monitoring; YOLOv8 and other algorithm porting technologies are affected by the underwater environment and have problems such as low recognition accuracy, easy false detection, and difficulty in real-time monitoring. Moreover, when adapted to underwater robots with low computing power, they face a sharp contradiction between computing power constraints and high accuracy requirements, which restricts the implementation of the technology.

[0005] The problems in combining general target detection algorithms with low-computing-power underwater robots are concentrated in three aspects: First, the YOLOv8 algorithm kernel (feature pyramid structure, dense prediction head) usually adopts a high computational density and high memory design, which makes it difficult to meet the real-time monitoring requirements of low-computing-power robots when directly combined with YOLOv8. Second, if existing low-computing-power adaptation schemes (pruning, quantization, etc.) are directly applied, they can significantly reduce detection accuracy and fail to match the actual needs of underwater monitoring. Third, there is currently no way to achieve a targeted combination with YOLOv8 by reasonably selecting and optimizing existing image enhancement algorithms, which makes it difficult to balance computational constraints, real-time performance and detection accuracy, and fails to fully leverage the mobile monitoring advantages of underwater robots.

[0006] In summary, there is an urgent need to develop a monitoring model construction method adapted to the underwater environment. By integrating a lightweight image enhancement algorithm with YOLOv8, efficient deployment of underwater robots with low computing power can be achieved, solving the dual problems of low computing power and high precision. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides an intelligent monitoring method and system for river and lake fish based on underwater robots. It integrates and coordinates the mobile patrol capabilities of underwater robots with the real-time intelligent recognition capabilities of deep learning algorithms to achieve automated, real-time, and high-precision in-situ monitoring and data analysis of fish targets in a large area of ​​natural rivers and lakes.

[0008] It includes two aspects, the first aspect:

[0009] A method for intelligent monitoring of river and lake fish based on underwater robots includes the following steps:

[0010] Step S1, Dataset Standardization and Preprocessing:

[0011] 1) Underwater robots were used to collect fish image data in the field;

[0012] 2) Select high-quality underwater fish images from publicly available academic datasets and authorized scientific research imagery to construct an auxiliary sample set;

[0013] 3) Perform multi-dimensional data augmentation preprocessing on the original fish image data;

[0014] Step S2, Construction of fish monitoring model framework: A basic fish monitoring model is constructed using the YOLOv8 framework;

[0015] Step S3, Model Structure Optimization:

[0016] 1) Introduce the lightweight and efficient local attention module ELA to enhance the edge contour features of the target;

[0017] 2) Introducing a large separable kernel attention module LSKA to capture the long-distance dependency between the fish body and the surrounding background, reducing the interference of background noise on contour recognition;

[0018] Step S4, Model Training Optimization:

[0019] The Shape-IoU loss function is introduced to replace the traditional bounding box CIoU regression loss, which precisely constrains the geometric alignment between the predicted box and the ground truth box;

[0020] Step S5, output the result:

[0021] After the model completes inference on the monitored images or video streams, it outputs multi-dimensional monitoring results.

[0022] Furthermore, in step S1, the data augmentation preprocessing employs an adaptive brightness transformation algorithm based on histogram adjustment, which performs linear perturbation within ±30% of the original image brightness to simulate the imaging effect under different time periods and changes in water transmittance, thereby enriching the illumination diversity of the image.

[0023] Furthermore, in step S3, the lightweight and efficient local attention module ELA is introduced. In the backbone network of the basic fish monitoring model constructed in the YOLOv8 framework, the standard C2f modules in layers 2, 4, 6, and 8 are replaced with an improved structure integrating ELA.

[0024] Furthermore, in step S3, the large separable kernel attention module LSKA is introduced. Before cross-layer feature fusion, the output feature maps of the 6th and 4th layers of the YOLOv8 shallow network are respectively introduced into LSKA. The features of the 6th layer after LSKA processing are concatenated to the 12th layer of the neck network, and the features of the 4th layer after LSKA processing are concatenated to the 16th layer of the neck network, realizing cross-layer fusion of enhanced features at different levels. The output feature maps of the 14th and 10th layers of the YOLOv8 shallow network are respectively introduced into LSKA. The features of the 14th layer after LSKA processing are concatenated to the 19th layer of the neck network, and the features of the 10th layer after LSKA processing are concatenated to the 22nd layer of the neck network.

[0025] Furthermore, the Shape-IoU loss function is calculated using the following formula:

[0026]

[0027] Where IoU is the predicted bounding box. With real frame The intersection and union ratio; The center coordinates of the prediction box The coordinates of the center of the true bounding box; The distance is Euclidean. for and The length of the diagonal of the minimum bounding rectangle; when t traverses the set , The width / height of the corresponding prediction box, The width / height of the corresponding actual bounding box; For adaptive scaling parameters, The corresponding width dimension , Time corresponds to height dimension ; These are the weighting coefficients for the shape loss term; The shape loss term is used to constrain the shape matching degree between the predicted bounding box and the ground truth bounding box. The calculation formula is as follows:

[0028]

[0029] Let be the angle between the predicted bounding box and the ground truth bounding box at the edge of the i-th segment; These are the contour feature vectors of the predicted bounding box and the ground truth bounding box.

[0030] Furthermore, the multi-dimensional monitoring results include: core target identification results, statistical analysis results, anomaly warning results, and data interface output.

[0031] The second aspect:

[0032] An intelligent monitoring system for river and lake fish based on an underwater robot includes: an underwater robot equipped with a high-definition optical camera and a fish monitoring system installed inside the underwater robot; the fish monitoring system includes:

[0033] Image acquisition module: used to acquire the image to be detected, and to perform data acquisition and preprocessing;

[0034] YOLOv8 detection module: connected to the image acquisition module, used to receive images and output target detection results;

[0035] ELA Enhancement Module: Connected to the YOLOv8 detection module, it is used to perform ELA enhancement operations on the image to be detected, perform local attention calibration on the edge details of the feature map, and enhance the edge contour features of the target.

[0036] LSKA Enhancement Module: Connected to the YOLOv8 detection module, it is used to perform LSKA enhancement operation on ELA enhanced images to capture long-distance correlation between fish and the background, and achieve accurate positioning in complex environments.

[0037] Result output module: Connected to the YOLOv8 detection module, used to output target detection results.

[0038] Storage module: Connected to the YOLOv8 detection module, used to store the image to be detected, the enhanced image, and the target detection results.

[0039] The beneficial effects of this invention are as follows: For underwater fish activity scenarios characterized by low contrast, high interference, and the need to balance global localization with local details, this invention embeds LSKA and ELA into the YOLOv8 backbone network and neck network, replacing the traditional lightweight attention module. ELA focuses on local areas, refining the target contour edges and details, and enhancing feature recognition. LSKA, with its wide-area feature coverage capability, filters background noise in dimly lit scenes, accurately locating the overall contour of fixed targets and avoiding missed detections. The introduction of the Shape-IoU loss function, through segmented edge direction alignment and contour feature similarity constraints, accurately adapts to the localization requirements of irregular fish shapes, significantly improving the fit between the detection box and the target. These advantages include:

[0040] 1. Improve monitoring accuracy and reliability: Effectively solve the problems of limited accuracy of traditional monitoring methods, easy false detection when general target detection algorithms are ported and applied, and insufficient accuracy of feature extraction. Through technical design adapted to special underwater environments, it reduces interference from factors such as water turbidity and uneven lighting, accurately identifies fish targets, and ensures the accuracy and effectiveness of monitoring data.

[0041] 2. Adapt to low-computing-power hardware scenarios: Select and optimize suitable image enhancement algorithms to reduce overall computational complexity and memory usage, achieve efficient adaptation with low-computing-power underwater robots, and ensure real-time monitoring requirements;

[0042] 3. Reduce ecological disturbance: Avoid the disturbance to the fish habitat caused by traditional methods such as manual fishing and statistics, adopt a gentler monitoring model, maintain the ecological balance of rivers and lakes, and help protect aquatic biodiversity.

[0043] 4. Enhance monitoring flexibility and coverage: Overcome the limitations of conventional fixed underwater video monitoring technology, which suffers from poor flexibility and narrow monitoring range. It can adapt to the migration patterns of fish school locations with seasonal and hydrological changes, reduce monitoring omissions and failures, and achieve more comprehensive coverage monitoring of rivers and lakes.

[0044] 5. Improve monitoring efficiency and reduce labor costs: Replace the inefficient traditional manual interpretation mode, reduce manpower input and time consumption through automated and intelligent technical solutions, avoid subjective errors, and support large-scale, long-term continuous monitoring needs.

[0045] The present invention will be explained in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0046] Figure 1 This is a flowchart of the intelligent monitoring method for river and lake fish based on underwater robots according to the present invention;

[0047] Figure 2 This is a framework diagram of the fish monitoring model of the present invention;

[0048] Figure 3 This is a structural diagram of the ELA module in the fish monitoring model of this invention;

[0049] Figure 4 This is a structural diagram of the LSKA module in the fish monitoring model of this invention. Detailed Implementation

[0050] Example 1:

[0051] This embodiment presents an intelligent monitoring method for river and lake fish based on an underwater robot, such as... Figure 1 As shown, it includes the following steps:

[0052] Step S1, Dataset Standardization and Preprocessing:

[0053] 1.1 Data Acquisition Method: Fish image data were collected in the field using an underwater robot (ROV / AUV) equipped with a high-definition optical camera; at the same time, a small number of high-quality underwater fish images were selected and included from publicly available academic datasets and authorized scientific research images to construct an auxiliary sample set.

[0054] 1.2 Data Acquisition: The data collection area covers various habitats, including typical river cross-sections, nearshore zones of reservoirs, and deep-water areas. To avoid the dispersal effect of underwater robots on fish schools and to reduce the Doppler effect, underwater robots were used for bottom-sitting monitoring, supplemented by image acquisition via a walking motion at an average speed of 0.3 m / s.

[0055] Finally, the collected data is returned to the starting point, ending the data collection mission. The captured video is then segmented into frames to obtain fish images.

[0056] 1.3 Standardization Processing:

[0057] 1) To improve the generalization ability of the training model and prevent overfitting, a systematic data augmentation preprocessing was performed on the original image. An adaptive brightness transformation algorithm based on histogram adjustment was adopted to perform linear perturbation within ±30% of the brightness of the original image, simulating the imaging effect under different time periods (such as strong light at noon and weak light at dusk) and changes in water transmittance, so as to enrich the illumination diversity of the image.

[0058] 2) By using batch processing scripts (Python, OpenCV) to randomly flip, rotate (angle range ±15°) and scale (scaling ratio range 0.8-1.2) images, the model's ability to recognize the invariance of target pose and scale is enhanced.

[0059] 3) Use the standardized annotation tool (LabelImg) to manually annotate the bounding boxes of fish targets in the images. The annotation specifications follow the PASCAL VOC format to ensure that each image generates a corresponding XML format annotation file.

[0060] Step S2, Construction of the fish monitoring model framework:

[0061] A basic fish monitoring model is built using the YOLOv8 framework, such as Figure 2 As shown, it mainly consists of the following four core components that work together:

[0062] (1) Input end: This module is responsible for data preprocessing and enhancement. In addition to basic image size normalization, it integrates Mosaic data enhancement (random four-image stitching), adaptive image scaling, and various color space transformation enhancement strategies to enrich data diversity in the early stage of training and improve the robustness of the model to scale and scene changes.

[0063] (2) Backbone: Its core consists of convolutional layers (conv), bottleneck layers (C2f), and a spatial pyramid pooling fast version (SPPF) module. The C2f module significantly reduces the number of computational parameters while enhancing feature reusability by segmenting the feature stream and fusing cross-stage information; the SPPF module achieves efficient fusion of receptive fields at different scales by concatenating multiple max pooling layers, thereby enhancing the model's ability to capture multi-scale targets.

[0064] (3) Neck Network: To enhance the representational power of the feature pyramid, the architecture combining the path aggregation network (PANet) and the feature pyramid network (FPN) was adopted and optimized to achieve efficient fusion of deep semantic features and shallow localization features.

[0065] (4) Head: The decoupled head design separates the classification task from the bounding box regression task, thereby obtaining more accurate category judgment and location regression results in complex scenarios.

[0066] YOLOv8 performs well in general target detection tasks. However, when applied directly to real-time underwater fish detection in rivers and lakes, it still faces a series of technical challenges due to the special operating environment: (1) Changes in water turbidity, uneven lighting, and color distortion result in low image contrast and blurred details, which seriously affect the accuracy of fish feature extraction; (2) Interference objects such as suspended particles and aquatic vegetation have visual similarity to fish targets, which can easily lead to false detections; (3) Existing offline processing modes cannot meet the real-time requirements.

[0067] Step S3, Model Structure Optimization:

[0068] (1) Introducing ELA:

[0069] ELA (Lightweight and Efficient Local Attention) refines the edge details of a fixed target (such as contour edges and local feature points), solving the problems of lost texture information and blurred contours in low-light scenes. It improves the Intersection over Union (IOU) between the bounding box and the target, reducing false detections. Its structure is as follows: Figure 3 As shown.

[0070] The specific implementation method is as follows: Figure 2 As shown, in the YOLOv8 backbone network, the standard C2f modules in layers 2, 4, 6, and 8 are replaced with an improved structure integrating ELA. The optimized structure table is as follows:

[0071] Table 1: Optimized Backbone Network Hierarchy

[0072]

[0073] Channel number matching: The number of input / output channels of the ELA module is consistent with that of the C2f module being replaced (e.g., in the YOLOv8 model, the number of channels of the second-layer C2f module is 64, and the number of channels of the corresponding ELA module is also set to 64).

[0074] Convolution kernel parameters: The ELA module uses a 3×3 convolution kernel with a stride of 1 and padding of 1 to ensure that the feature map size remains unchanged; the GELU activation function is used to improve gradient flow efficiency.

[0075] Integration method: The ELA module is embedded in the cross-stage branch of the C2f module, replacing the dense convolutional layers of the original branch, while retaining the residual connection structure of the C2f module to ensure feature reusability.

[0076] ELA employs a multi-branch lightweight convolutional + short connection design. Compared to the cross-stage dense convolution of C2f, it reduces redundant channel interaction calculations while maintaining feature representation capabilities. Especially in C2f layers with many repetitions (such as layers 4 / 6, which are repeated 6 times), the computational cost is significantly reduced after replacement, with a reduction of approximately 8%-15% in tests.

[0077] (2) Introducing LSKA:

[0078] LSKA (Large Separable Kernel Attention) achieves wide-area feature coverage through a separable large kernel and dilated convolution. It can capture the long-distance dependency between the fish body and the surrounding background without significantly increasing the computational cost, effectively distinguishing the fish outline from interference objects such as suspended particles and aquatic vegetation, and reducing the interference of background noise on outline recognition.

[0079] LSKA module such as Figure 4 As shown, this module is a computationally efficient visual attention mechanism. Its core lies in decomposing the traditional two-dimensional large kernel convolution into cascaded one-dimensional depthwise convolution and one-dimensional dilated convolution, thereby significantly reducing the number of parameters and computational complexity while maintaining a large receptive field.

[0080] Before performing cross-layer feature fusion, such as Figure 2 As shown, LSKA is introduced into the output feature maps of layers 6 and 4 of the YOLOv8 shallow network. The features of layer 6 after LSKA processing are concatenated to layer 12 of the neck (first feature concatenation layer), and the features of layer 4 after LSKA processing are concatenated to layer 16 of the neck (secondary feature concatenation layer), realizing cross-layer fusion of enhanced features at different levels. LSKA is also introduced into the output feature maps of layers 14 and 10 of the YOLOv8 shallow network. The features of layer 14 after LSKA processing are concatenated to layer 19 of the neck, and the features of layer 10 after LSKA processing are concatenated to layer 22 of the neck.

[0081] LSKA can simulate large receptive field visual attention using decomposable large convolutional kernels, thus efficiently capturing a wide range of contextual dependencies and spatial structure information in images, while having relatively low computational overhead. In the YOLOv8 network, feature maps become smaller in size, have more channels, and stronger semantic information as the network deepens. Layers 10 and 14 correspond to moderately shallow features with small receptive fields, mainly focusing on local details. Introducing LSKA can inject a wide range of contextual information into these detailed features. Layers 19 and 22 typically correspond to key feature maps fused from different detection heads. Inputting the LSKA-enhanced feature information into the fusion nodes of the neck network (layers 19 and 22) not only fuses the original detailed and semantic features but also incorporates features that have been modulated and enhanced with global contextual information, improving detection accuracy.

[0082] Table 2: Optimized Neck Network Hierarchy Table

[0083]

[0084] Fish monitoring models address core challenges in underwater robots (ROV / AUV): complex underwater environments require long-range feature capture, while edge devices have limited computing power. LSKA's design is highly suited to this scenario, as detailed below:

[0085] From a technical perspective, LSKA is a lightweight and optimized version of the standard Large Kernel Attention (LKA). Its core principle is to decompose a two-dimensional large kernel convolution (e.g., 21×21) into a series of "n×1 vertical + 1×n horizontal" one-dimensional convolutions, where n is the preset kernel size. This decomposition reduces parameter complexity from... Reduced to (k is the core size, C is the number of channels) It is perfectly adapted to the computing power limitations of embedded low-power chips (such as Jetson Xavier NX) used in underwater robots, and can ensure real-time detection (inference latency ≤50ms) without consuming too many resources.

[0086] The key advantage lies in the fact that LSKA reduces costs without sacrificing efficiency, fully retaining its long-distance spatial dependence capture capability, and accurately matching underwater detection needs. Underwater fish are easily obscured by aquatic plants and suspended particles, and high water turbidity and uneven lighting result in low target-background contrast. Traditional local attention modules are prone to missed detections and false detections due to their limited receptive field; while LSKA covers a larger space through one-dimensional large kernel aggregation, which can capture the long-distance correlation between fish and the background, achieving accurate positioning in complex environments, and is especially suitable for the detection of schooling fish and low-contrast targets in low light.

[0087] From the perspective of attention generation logic, LSKA can specifically enhance the effective features of underwater targets and suppress background noise (seaweed texture, particle reflection) by decomposing convolutions to extract context, generating weights channel by channel, and calibrating features element by element, thereby improving the signal-to-noise ratio of feature maps. This capability can be seamlessly integrated with YOLOv8 models without the need for additional preprocessing modules, simplifying the system architecture.

[0088] LSKA combines the advantages of a large receptive field with a lightweight design, effectively addressing the challenges of limited computing power and difficult underwater environment detection in underwater robots. It is the optimal attention module choice for this scenario.

[0089] This design aims to maintain the network's deep feature extraction capabilities while adaptively selecting important features and suppressing redundant responses through a local attention mechanism, thereby reducing model complexity while improving the discriminative power of feature representation. The ELA module employs a lightweight dual-path attention mechanism, extracting spatial and channel features in parallel through grouped convolutions and depthwise separable convolutions, and generating collaborative attention weights to dynamically recalibrate the feature maps. This design significantly enhances the model's ability to focus on key features of underwater fish with almost no increase in the number of parameters, while effectively suppressing background interference and feature redundancy.

[0090] Step S4, Model Training Optimization:

[0091] The optimized model was used for end-to-end training on the collected underwater fish dataset. During the initial training phase, the input data underwent a unified preprocessing procedure: first, all training images were reshaped to a resolution of 640×640 pixels using bilinear interpolation, and then normalized (e.g., Min-Max) to adapt to the model's fixed input size requirements.

[0092] In terms of model optimization strategy, the Shape-IoU loss function (Shape-aware Intersection over Union Loss) is introduced to replace the traditional bounding box CIoU regression loss, which precisely constrains the geometric alignment between the predicted box and the ground truth box. The calculation formula is as follows:

[0093]

[0094] Where IoU is the predicted bounding box. With real frame The intersection and union ratio; The center coordinates of the prediction box The coordinates of the center of the true bounding box; The distance is Euclidean. for and The length of the diagonal of the minimum bounding rectangle; when t traverses the set , The width / height of the corresponding prediction box, The width / height of the corresponding actual bounding box; For adaptive scaling parameters, The corresponding width dimension , Time corresponds to height dimension ; Width dimension ∈ [0.15, 0.25], height dimension ∈[0.15,0.25]; This is the weighting coefficient for the shape loss term, with a value ranging from 0.3 to 0.5; The shape loss term is used to constrain the shape matching degree between the predicted bounding box and the ground truth bounding box. The calculation formula is as follows:

[0095]

[0096] Let be the angle between the predicted bounding box and the ground truth bounding box at the edge of the i-th segment; These are the contour feature vectors of the predicted bounding box and the ground truth bounding box.

[0097] Compared to traditional IoU loss, which only optimizes the distance between overlapping regions and the center, Shape-IoU, through segmented directional alignment loss and contour feature similarity loss, forces the model to learn the geometric features of the target. This results in stronger modeling capabilities for irregular shapes such as bending and tilting of underwater fish, significantly reducing localization errors in complex backgrounds. During training, the loss curve decreases; training ends when the loss curve is essentially parallel to the X-axis, indicating good training results. The model weights for this dataset can then be obtained after training.

[0098] Step S5, output the result:

[0099] After the model completes inference on the monitored images / video streams, it automatically outputs multi-dimensional monitoring results, including:

[0100] 1. Core results of target recognition: Real-time output of the category (such as specific fish species name), quantity, confidence level (probability value) and precise bounding box coordinates of fish targets in image / video frames. The annotation results can be superimposed on the original image / video stream to form a visual monitoring screen;

[0101] 2. Statistical analysis results: Summarize monitoring data according to preset time periods (such as hourly or daily) and generate statistical reports such as fish population change trends, proportion of dominant fish species, and fish distribution density in key monitoring areas (such as spawning grounds and migration channels);

[0102] 3. Abnormal early warning results: When abnormal situations such as sudden changes in fish population, the appearance of rare and protected fish species, or illegal fishing are detected, an early warning signal is automatically triggered, and key information such as the time, location, and image evidence of the abnormal event is output simultaneously;

[0103] 4. Data Interface Output: Supports the synchronization of raw identification results, statistical data and early warning information to the back-end monitoring and management platform, mobile terminal or local storage module of low computing power underwater robot through standardized data interface, so that staff can view in real time, conduct subsequent data analysis and refer to fisheries management decisions, while retaining complete monitoring logs to ensure data traceability.

[0104] This invention addresses the technical problems of low efficiency and significant ecological interference in traditional methods for monitoring fish in rivers and lakes, as well as the poor adaptability of general target detection algorithms and the prominent contradiction between low computing power and high accuracy. It adopts an integrated technical solution combining underwater robot mobile data collection, lightweight attention enhancement, and shape perception loss constraints. The innovations are as follows:

[0105] Model structure co-optimization: The ELA local attention module and the LSKA long-range dependency capture module are co-embedded in the YOLOv8 network. To address the problems of local blurring and global noise, an ELA+LSKA co-engineering architecture is designed. ELA focuses on local edge details to solve the problem of fish outline blurring; LSKA captures long-range dependencies to distinguish fish from background interference. The two functions complement each other rather than being simply superimposed.

[0106] Shape-aware loss constraint: To address the localization error problem of irregular fish shapes, the Shape-IoU loss function is introduced. By constraining the segmented edge direction angles and contour feature vectors, more accurate bounding box regression than CIoU is achieved, reducing localization error by more than 60%.

[0107] Low-computing-power adaptation design: Both ELA and LSKA are lightweight modules. After integration, the number of model parameters only increases by 12.5%, and the inference latency is controlled within 40ms. They are perfectly adapted to the embedded hardware of underwater robots, solving the contradiction between low computing power and high precision.

[0108] Actual testing and verification have shown that this invention successfully identified 12 species of river and lake fish during 30 days of continuous monitoring in a 50 km² area of ​​a large reservoir, with a population counting error of ≤3% and an accuracy rate of 95.8% for identifying rare fish. Compared with traditional manual monitoring, efficiency is increased by 8 times, labor costs are reduced by 90%, and there is no ecological interference throughout the process. Even in complex underwater environments with turbidity >50 NTU, it still maintains an average detection accuracy of 73.2%, far exceeding existing similar technologies. This invention achieves a technological breakthrough in automated, high-precision, low-interference, and wide-coverage river and lake fish monitoring, providing efficient and reliable technical support for aquatic ecological protection and fishery resource management.

[0109] Example 2:

[0110] This invention provides an intelligent monitoring system for river and lake fish based on an underwater robot, used to perform the aforementioned target detection method. The system includes: an R1-45 type underwater robot (ROV / AUV) equipped with a high-definition optical camera and a fish monitoring system deployed in the embedded hardware of the underwater robot; the fish monitoring system includes:

[0111] 1. Image acquisition module: Used to acquire the image to be detected, perform data collection and preprocessing;

[0112] 2. YOLOv8 detection module: Connected to the image acquisition module, used to receive images and output target detection results.

[0113] 3. ELA Enhancement Module: Connected to the YOLOv8 detection module, it is used to perform ELA enhancement on the image to be detected, perform local attention calibration on the edge details of the feature map, and enhance the edge contour features of the target;

[0114] 4. LSKA Enhancement Module: Connected to the YOLOv8 detection module, it is used to perform LSKA enhancement operation on ELA enhanced images. It can capture long-distance correlation between fish and background, and achieve accurate positioning in complex environments. It is especially suitable for the detection of fish in groups and low-contrast targets in low light.

[0115] 5. Result Output Module: Connected to the YOLOv8 detection module, used to output target detection results.

[0116] 6. Storage module: Connected to the YOLOv8 detection module, used to store the image to be detected, the enhanced image, and the target detection results.

[0117] The above description is only used to illustrate the technical solutions of the present invention and is not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention (such as the application of various formulas, the order of steps, etc.) without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A river and lake fish intelligent monitoring method based on an underwater robot, characterized in that, Includes the following steps: Fish image data were collected in the field using an underwater robot; high-quality underwater fish images were selected from publicly available academic datasets and authorized scientific research image data to construct an auxiliary sample set; and multi-dimensional data augmentation preprocessing was performed on the original fish image data. A basic fish monitoring model was built using the YOLOv8 framework; A lightweight and efficient local attention module, ELA, is introduced to enhance the edge contour features of the target. A large separable kernel attention module LSKA is introduced to capture the long-distance dependency between the fish body and the surrounding background, reducing the interference of background noise on contour recognition; The Shape-IoU loss function is introduced to replace the traditional bounding box CIoU regression loss, which constrains the geometric alignment between the predicted box and the ground truth box. After the model completes inference on the monitored images or video streams, it outputs multi-dimensional monitoring results. 2.The method according to claim 1, wherein, The data augmentation preprocessing employs an adaptive brightness transformation algorithm based on histogram adjustment, which linearly perturbs the brightness of the original image within ±30% to simulate the imaging effect under different time periods and changes in water transmittance, thereby enriching the illumination diversity of the image. 3.The method according to claim 1, wherein, The lightweight and efficient local attention module ELA is introduced, and in the backbone network of the basic fish monitoring model built in the YOLOv8 framework, the standard C2f modules in layers 2, 4, 6, and 8 are replaced with an improved structure that integrates ELA. 4.The intelligent monitoring method of river and lake fish based on underwater robots according to claim 1, characterized in that, The large separable kernel attention module LSKA is introduced. Before cross-layer feature fusion, the output feature maps of the 6th and 4th layers of the YOLOv8 shallow network are respectively introduced into LSKA. The features of the 6th layer after LSKA processing are concatenated to the 12th layer of the neck network, and the features of the 4th layer after LSKA processing are concatenated to the 16th layer of the neck network, realizing cross-layer fusion of enhanced features at different levels. The output feature maps of the 14th and 10th layers of the YOLOv8 shallow network are respectively introduced into LSKA. The features of the 14th layer after LSKA processing are concatenated to the 19th layer of the neck network, and the features of the 10th layer after LSKA processing are concatenated to the 22nd layer of the neck network. 5.The intelligent monitoring method of river and lake fish based on underwater robot according to claim 1, characterized in that, The Shape-IoU loss function is calculated using the following formula: Where IoU is the predicted bounding box. With real frame The intersection and union ratio; The center coordinates of the prediction box The coordinates of the center of the true bounding box; The distance is Euclidean. for and The length of the diagonal of the minimum bounding rectangle; when t traverses the set , The width and height of the corresponding prediction box, Corresponding to the width and height of the actual bounding box; For adaptive scaling parameters, The corresponding width dimension , Time corresponds to height dimension ; These are the weighting coefficients for the shape loss term; The shape loss term is used to constrain the shape matching degree between the predicted bounding box and the ground truth bounding box. The calculation formula is as follows: is the direction angle between the predicted bounding box and the ground truth bounding box on the i-th segment edge; is the contour feature vector of the predicted bounding box and the ground truth bounding box.

6. The intelligent monitoring method for river and lake fish based on underwater robots according to claim 1, characterized in that, The multi-dimensional monitoring results include: core target identification results, statistical analysis results, anomaly warning results, and data interface output.

7. A smart monitoring system for river and lake fish based on an underwater robot, as described in any one of claims 1-6, characterized in that, include: An underwater robot equipped with a high-definition optical camera and a fish monitoring system deployed in the embedded hardware of the underwater robot; The fish monitoring system includes: Image acquisition module: used to acquire the image to be detected, and to perform data acquisition and preprocessing; YOLOv8 detection module: connected to the image acquisition module, used to receive images and output target detection results; ELA Enhancement Module: Connected to the YOLOv8 detection module, it is used to perform ELA enhancement operations on the image to be detected, perform local attention calibration on the edge details of the feature map, and enhance the edge contour features of the target. LSKA Enhancement Module: Connected to the YOLOv8 detection module, it is used to perform LSKA enhancement operation on ELA enhanced images to capture long-distance correlation between fish and the background, and achieve accurate positioning in complex environments; Result output module: connected to the YOLOv8 detection module, used to output target detection results; Storage module: Connected to the YOLOv8 detection module, used to store the image to be detected, the enhanced image, and the target detection results.

Images