Rail type fish condition intelligent inspection robot control system

Abstract

The invention provides a rail type fish condition intelligent inspection robot control system which comprises a factory factory, M fish ponds are arranged in the factory factory, and M is a positive integer larger than or equal to 1; an inspection track for the inspection robot to move is arranged on the factory building, and the inspection robot is arranged on the track; the system further comprises an industrial personal computer and an inspection platform. After the inspection robot moves to a target fishpond, a fishpond water surface picture is shot, the shot picture is uploaded to the edge industrial personal computer, and the industrial personal computer recognizes the number of dead fishes and the number of fish floating heads in the picture and uploads a result to the digital twin management and control cloud platform. According to the invention, the number of dead fishes and fish floating heads in the fishpond picture can be identified, and the labor cost is saved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligence, in particular to a track type fish emotion intelligent inspection robot control system. BACKGROUND

[0002] With the development of social economy and urbanization, the global arable land area is decreasing, and the facility planting industry is developing rapidly. Greenhouses create a microclimate suitable for crop growth and aquaculture, and monitoring is crucial for efficient and stable implementation of planting and breeding tasks. Initially, greenhouse inspection was conducted by manual labor, which was low-cost but prone to misjudgment due to the influence of staff experience. Later, through the installation of sensors to monitor environmental indicators, efficiency and intensity were optimized, but the monitoring system lacked flexibility and long-term operation capability, and true automation was not achieved. SUMMARY

[0003] The present application aims to at least solve the technical problems existing in the prior art, and particularly innovatively provides a track type fish emotion intelligent inspection robot control system.

[0004] In order to achieve the above-mentioned purpose of the present application, the present application provides a track type fish emotion intelligent inspection robot control system, which comprises a factory building, M fish ponds are arranged in the factory building, and M is a positive integer greater than or equal to 1;

[0005] A inspection track for the movement of the inspection robot is arranged on the factory building, and the inspection robot is arranged on the track;

[0006] Further comprising an industrial computer and an inspection platform;

[0007] After the inspection robot moves to the target fish pond, it takes a picture of the water surface of the fish pond, uploads the taken picture to the edge industrial computer, identifies the number of dead fish and fish floating heads in the picture by the industrial computer, and uploads the result to the digital twin management and control cloud platform.

[0008] In a preferred embodiment of the present application, if the number of dead fish exceeds the set number, an early warning information is pushed to the WeChat public number bound to the system for timely reminding, and the early warning information is that the number of dead fish exceeds the set number;

[0009] Or / and set the daily inspection frequency, inspection photographing point and charging time of the robot one or any combination;

[0010] Or / and M is 56, there are 7 fish ponds in one row and 8 fish ponds in one column, and there are a total of 56 fish ponds;

[0011] Or / and the inspection track is a ring-shaped distributed track;

[0012] Or / and the inspection robot is an inspection robot carrying a monitoring camera.

[0013] The application also discloses a control method of the track fish emotion intelligent inspection robot.

[0014] S1, a monitoring camera acquires a fish pond water surface monitoring video;

[0015] S2, the video is transmitted to an industrial computer, and a water surface picture is obtained by frame cutting of the industrial computer;

[0016] S3, an image containing a fish floating head or / and a dead fish is selected from the picture, and the image is respectively taken as a training set, a test set and a verification set in proportion;

[0017] S4, the images in the training set are introduced into a YOLOv12 model to train the YOLOv12 model, the images in the test set are introduced into the trained YOLOv12 model to test the trained YOLOv12 model, and the images in the verification set are introduced into the tested YOLOv12 model to verify the tested YOLOv12 model; finally, a target YOLOv12 model is obtained; the YOLOv12 model can be an existing YOLOv12 model.

[0018] S5, a target fish pond picture photographed by the inspection robot is introduced into the target YOLOv12 model, and an output result is obtained.

[0019] In a preferred embodiment of the application, in step S1, the fish ponds are fish ponds under different water quality and light conditions;

[0020] Or / and in step S3, the proportions of the training set, the test set and the verification set are 8:1:1;

[0021] Or / and in step S3, a LabelImg software is used for accurate marking on the fish floating head or / and the dead fish area, and the whole fish is not marked, because the floating head behavior is determined by the characteristics of the head exposed to the water surface, and the relative position of the fish head on the water surface is stable and less disturbed by water waves, so as to facilitate accurate positioning and detection of the model; through this targeted marking and scene adaptive acquisition mode, combined with the multi-scale detection capability of the model, efficient identification and real-time monitoring of the floating head and the dead fish can be realized under complex water surface conditions.

[0022] In a preferred embodiment of the application, in step S4, the YOLOv12 model comprises a YOLOv12 backbone network, a YOLOv12 neck network and a YOLOv12 detection head;

[0023] A data output end of the YOLOv12 backbone network is connected with a data input end of the YOLOv12 neck network, and a data output end of the YOLOv12 neck network is connected with a data input end of the YOLOv12 detection head;

[0024] The YOLOv12 backbone network is responsible for extracting multi-level feature information from the input image, providing a basic feature representation for subsequent target detection;

[0025] The YOLOv12 neck network is used to fuse and enhance the multi-scale features extracted by the backbone network, so as to detect targets of different sizes;

[0026] The YOLOv12 detection head is used to convert the multi-scale fusion features output by the neck network into the final detection results.

[0027] In a preferred embodiment of the present application, an improved CSPDarknet structure is included in the YOLOv12 backbone network, which includes dividing the feature flow into two parts by cross-stage partial connection, one part is directly passed to the next stage, and the other part is combined with the output of the next stage through a jump connection; in this way, the gradient flow in the network is effectively separated, avoiding the gradient vanishing problem that may occur in traditional deep networks, while reducing the computational load of each stage; in addition, in the CSPDarknet, the feature flow is effectively separated, so that each part of the network focuses on learning features of different levels; this not only improves the feature expression ability of the model, but also enhances the flexibility of the network in handling complex tasks; through this separation, the CSPDarknet maintains high computational efficiency at each stage; the CSPDarknet integrates a RepConv convolution module and a SPPF module; RepConv is integrated into the convolution layer to improve feature extraction performance, and in deep networks, it can significantly improve the learning ability of the network; SPPF is used to enhance the receptive field of the network, which can capture more rich spatial information while maintaining computational efficiency;

[0028] The YOLOv12 neck network adopts an improved PAN-FPN structure, which effectively combines high-level semantic features and low-level fine-grained features, retaining the detailed information of small targets while enhancing the semantic expression of large targets; among them, FPN is responsible for expanding high-level features to low-level through upsampling, while enhancing the semantic expression of low-level features through downsampling, and PAN optimizes the information flow of low-level features, ensuring that low-level detailed information can be efficiently passed to deeper network levels; after the fusion of the two paths, PAN-FPN can combine the semantic information of high-level and the fine-grained information of low-level, effectively improving the accuracy in target detection tasks;

[0029] PAN-FPN also introduces a C2f lightweight feature processing module, which improves information flow efficiency and reduces the risk of gradient vanishing through cross-layer connection and feature reuse;

[0030] The YOLOv12 detection head includes a multiscale detection strategy of the YOLO series, and prediction is respectively performed on feature maps of different resolutions, while the detection requirements of large, medium and small targets of different sizes are considered; each scale prediction branch outputs a tensor with a fixed number of channels through a convolution layer, and the tensor contains the center coordinates, width and height of the bounding box, target confidence and class probability;

[0031] The YOLOv12 detection head uses an Anchor-Free mechanism in the decoding process to simplify anchor matching and improve the detection accuracy of small targets and dense scenes;

[0032] The YOLOv12 detection head cooperates with a loss function, the loss function includes one or any combination of CIoU, DFL and classification cross-entropy, and the balance between detection performance and inference speed is realized through end-to-end optimization, thereby providing a guarantee for real-time high-precision target detection.

[0033] In a preferred embodiment of the present application, the YOLOv12 weight pre-trained on the COCO large data set is loaded in step S3, the model can be fine-tuned on the basis of the existing general features, and is specially adapted to the detection task of fish floating heads and dead fish; this transfer learning method shortens the training time and improves the detection accuracy and generalization ability.

[0034] In a preferred embodiment of the present application, the learning rate adopts a dynamic learning rate strategy in stages in step S3, which can avoid the model from falling into a local optimum and ensure stable convergence;

[0035] The optimizer selects SGD, and is flexibly adjusted according to the data size and convergence speed requirement;

[0036] In the inference stage, the confidence threshold and the NMS threshold are optimized; a lower threshold helps to improve the recall rate, and a higher threshold can reduce false positives.

[0037] The present application also discloses a computer system, comprising:

[0038] A processor;

[0039] A memory for storing processor executable instructions;

[0040] The processor is configured to implement the track type fish emotion intelligent inspection robot control method when the executable instructions are executed.

[0041] The present application also discloses a computer readable storage medium, comprising:

[0042] A memory having a computer program stored thereon;

[0043] A processor is configured to execute the program in the memory to realize the track fish intelligent inspection robot control method.

[0044] Therefore, the present application can identify the number of dead fish and fish floating heads in the fish pond picture, and save labor costs.

[0045] Additional aspects and advantages of the present application will be made apparent by the following description of embodiments thereof, given as a non-restrictive example, with reference to the annexed drawings. BRIEF DESCRIPTION OF DRAWINGS

[0046] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood by considering the following detailed description, from which the aspects and advantages of the present application will be readily understood, taken in conjunction with the accompanying drawings.

[0047] Figure 1 is a flow connection diagram of the present application.

[0048] Figure 2 is a track fish inspection display diagram of the present application.

[0049] Figure 3 is a track robot physical display diagram of the present application.

[0050] Figure 4 is a fish floating head image display diagram of the present application.

[0051] Figure 5 is a YOLOv12 structure diagram of the present application.

[0052] Figure 6 is a track inspection platform display diagram of the present application. DETAILED DESCRIPTION

[0053] The embodiments of the present application are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application.

[0054] The track fish intelligent inspection robot system adopts a "track mobile platform + edge computing + cloud cooperation" architecture, covers multiple (7x8) fish ponds through a ring distributed track, carries a monitoring camera to realize full-process unmanned inspection of high-end varieties such as sea bass, mandarin fish and hybrid sturgeon, such as Figure 2The system builds a robot inspection track. The inspection platform can set the daily inspection frequency, inspection photographing point, and charging time. The track inspection robot takes pictures of the fish pond surface after reaching the target fish pond. The pictures are uploaded to the edge industrial computer through WiFi. The industrial computer deploys an improved yolov12 fish condition analysis model to identify the number of dead fish and fish floating heads in the pictures and upload the results to the digital twin management and control cloud platform. The cloud platform sets an early warning rule. If the number exceeds the set number, the early warning information will be pushed to the WeChat public number bound to the system. The fish condition inspection robot can timely discover breeding disasters through efficient model analysis and early warning mechanism, avoid large-scale fish death, and reduce the economic loss of breeders.

[0055] The inspection robot adopts a native inspection robot, such as Figure 3 The step motor can be set to an S-shaped curve to realize intelligent start-stop buffering, and the maximum moving speed is 20 m / min. The inspection point can be set through teaching, and the inspection plan list is generated to execute the inspection task at a fixed time. After the inspection is photographed, the inspection pictures are transmitted through WiFi.

[0056] The inspection cloud platform also includes robot state information, remote control, preset position setting, and inspection task setting, as shown in Figure 6

[0057] Robot state information: The robot state information can display the idle, charging, inspection, and maintenance states of the robot in real time.

[0058] Remote control:

[0059] 1) Monitoring control. The photographing, video recording, and light switching can be controlled.

[0060] 2) Robot control. The robot can be controlled to move left, right, up, down, and stop.

[0061] 3) Gimbal control. The gimbal can be controlled to tilt up, tilt down, turn left, turn right, tilt up and turn left, tilt up and turn right, tilt down and turn left, tilt down and turn right, automatic scanning, focus +, focus -, focus +, focus -, aperture +, aperture, photographing, video recording, and lighting.

[0062] 4) Preset position setting. After the robot reaches the monitoring location, click new to set the inspection requirements according to the actual scene.

[0063] 5) Inspection task strategy. The robot can be set to execute the inspection task at a certain time, frequency, and cycle to set an inspection plan for the robot.

[0064] (1) Main equipment

[0065] ​Monitor camera (resolution 4MP), training graphics card (RTX3060), win11 operating system.

[0066] (2) Test site

[0067] Chongqing Academy of Agricultural Sciences Agricultural High-tech Park Fish and Vegetable Symbiosis AI Factory.

[0068] (3) Test materials

[0069] In the video data acquisition stage, the continuous monitoring video is frame extracted at the rate of one frame per second to generate static image data, and 13591 fish floating head and dead fish images are randomly selected from the factory circulating water aquaculture system under different water quality and light conditions, as shown in Figure 4 The allocation ratio of 8-1-1 (i.e. 80%, 10%, 10%) is used to randomly allocate pictures as training set, test set and validation set.

[0070] (4) Data annotation

[0071] In the data annotation process, the head region of the floating head fish is accurately marked by using LabelImg software, and the whole fish is not labeled. The reason is that the floating head behavior is mainly judged by the characteristics of the head exposed to the water surface, and the relative position of the fish head on the water surface is stable and less disturbed by water waves, which is convenient for the model to accurately locate and detect. Through this targeted annotation and scene adaptation collection method, combined with the multi-scale detection ability of the model, efficient identification and real-time monitoring of floating head and dead fish can be realized under complex water surface conditions.

[0072] (5) Algorithm principle

[0073] The present application uses the current YOLOv12 model to monitor the floating head and dead fish of fish. The overall framework of the YOLOv12 model is shown in Figure 5

[0074] (1) Backbone network

[0075] ​The backbone network of YOLOv12 is mainly responsible for extracting multi-level feature information from the input image, providing a foundation for subsequent target detection. This part is usually composed of an improved CSPDarknet structure, which divides the feature flow into two parts through Cross Stage Partial Connections (CSP). One part is directly passed to the next stage, and the other part is combined with the output of the next stage through a skip connection. In this way, the gradient flow in the network is effectively separated, avoiding the gradient vanishing problem that may occur in traditional deep networks, while reducing the computational load of each stage. In addition, in the CSPDarknet, the feature flow is effectively separated, so that each part of the network focuses on learning different levels of features. This not only improves the feature representation ability of the model, but also enhances the flexibility of the network in handling complex tasks. Through this separation, the CSPDarknet maintains a high computational efficiency at each stage. More importantly, the CSPDarknet integrates the RepConv convolution module and the SPPF module (Spatial Pyramid Pooling-Fast). RepConv is integrated into the convolution layer to improve feature extraction performance, especially in deep networks, which can significantly improve the learning ability of the network. SPPF is used to enhance the receptive field of the network, enabling it to capture more rich spatial information while maintaining computational efficiency, especially the combination of global and local features.

[0076] (2) neck

[0077] The neck network of YOLOv12 is located between the backbone network and the detection head. Its main function is to fuse and enhance the multi-scale features extracted by the backbone network, so as to detect targets of different sizes. This part usually uses an improved PAN-FPN (Path Aggregation Network combined with Feature Pyramid Network) structure to effectively combine high-level semantic features with low-level fine-grained features, retaining the details of small targets and enhancing the semantic expression of large targets. Specifically, FPN is responsible for expanding high-level features to low-level features through upsampling, while enhancing the semantic expression of low-level features through downsampling. PAN further optimizes the information flow of low-level features, ensuring that low-level details can be efficiently transmitted to deeper network levels. After the fusion of the two paths, PAN-FPN can combine the semantic information of high-level features with the fine-grained information of low-level features, effectively improving the accuracy in the target detection task. PAN-FPN also introduces lightweight feature processing modules such as C2f, which improve information flow efficiency and reduce the risk of gradient vanishing through cross-layer connections and feature reuse. Finally, the neck sends the fused multi-scale features to the prediction head (Prediction Head) to achieve accurate detection and positioning of fish targets of different scales.

[0078] (3) Detection head

[0079] The detection head of YOLOv12 is responsible for converting the multi-scale fusion features output by the neck network into the final detection results, including the class, position, and confidence of the target. This detection head continues the multi-scale detection strategy of the YOLO series, predicting on feature maps of different resolutions to simultaneously meet the detection needs of large, medium, and small targets of different sizes. Each scale prediction branch outputs a fixed number of tensors through convolution layers, which contain the center coordinates of the bounding box, width and height, target confidence, and class probabilities. The detection head of YOLOv12 uses the Anchor-Free mechanism during decoding to simplify anchor matching and improve the detection accuracy of small targets and dense scenes. In addition, this detection head cooperates with loss functions such as CIoU, DFL, and classification cross-entropy to balance detection performance and inference speed through end-to-end optimization, providing a guarantee for real-time high-precision target detection.

[0080] (6) Model training and optimization and results

[0081] Model training and optimization mainly includes model pre-training, model transfer training, and parameter optimization stages.

[0082] a. Model pre-training and transfer training

[0083] In the application of factory fish floating head and dead fish detection, there are two problems in directly training the YOLOv12 model from scratch: first, the limited scale of data from the farm may lead to model overfitting; second, the complex water environment makes it difficult to learn stable feature patterns with small samples. By loading the pre-trained YOLOv12 weights on large datasets such as COCO, the model can fine-tune on the basis of existing general features to specifically adapt to the detection task of fish floating head and dead fish. This transfer learning method significantly shortens the training time while improving the accuracy and generalization ability of detection.

[0084] b. Parameter tuning

[0085] In terms of learning rate and optimizer, a phased dynamic learning rate strategy (such as cosine annealing or OneCycleLR) can be used to avoid the model from falling into a local optimum and ensure stable convergence; the optimizer is usually SGD, which can be adjusted flexibly according to the data size and convergence speed requirements. In the inference stage, the confidence threshold and NMS (Non-Maximum Suppression) threshold are tuned. Lower threshold helps to improve recall rate, and higher threshold can reduce false positives. Combined with the actual application of factory farming, the threshold can be dynamically set.

[0086] c. Training results

[0087] To comprehensively verify the overall performance of the model, the five common weight models (n, s, m, l, and x) of YOLOv12 are trained respectively, and the final detection results are shown in Table 1.

[0088] Table 1 Comparison of different YOLOv12 weight results

[0089]

[0090] The experimental results show that with the gradual increase of the model size from YOLOv12n to YOLOv12x, the detection accuracy and comprehensive performance show a steady upward trend. The lightweight YOLOv12n model achieves 84.9%, 69.4% and 77.6% in accuracy, regression and m AP50, respectively, which is suitable for scenarios with high real-time requirements and limited computing power. With the increase of network depth and parameter quantity, YOLOv12m and YOLOv12l exceed 67% in accuracy and mAP50-95, achieving a balance between detection accuracy and computational overhead. The highest performance YOLOv12x model achieves 92.5% and 80.6% in accuracy and regression, respectively, with mAP50 and mAP50-95 of 87.8% and 71.0%, respectively, showing the best overall detection effect and being suitable for fish floating head and dead fish detection tasks with high accuracy requirements. Overall, the YOLOv12 series model can be flexibly selected according to different hardware conditions and application requirements, achieving robust detection capability in factory farming environment.

[0091] (7) Model deployment

[0092] The deployment environment of the model is Windows 10 operating system, and the hardware platform uses NVIDIA RTX 3050 graphics card. During deployment, first configure the CUDA and cuDNN libraries compatible with the graphics card driver, and build the running environment based on Python and PyTorch framework to fully utilize the parallel computing capability of GPU. The trained weight file is converted to ONNX format, and combined with TensorRT acceleration in the inference stage to significantly reduce the delay and achieve real-time detection of 30-60 frames per second. The entire deployment process includes video stream acquisition, image preprocessing, model inference and result post-processing, and the detection results are quickly visualized through GPU hardware decoding. On this basis, the YOLOv12 model can stably run in complex water surface environment and meet the real-time monitoring requirements of fish floating head in factory farming scenarios.

[0093] Through experimental data and parameter analysis, the final model can effectively identify and detect the number of fish floating head and dead fish, with high real-time performance and the condition of landing deployment. At the same time, through continuous iteration of test set quantity and model retraining in the later stage, there is room for further improvement of the detection accuracy and precision of the model.

[0094] The patent proposes a track-type fish emotion intelligent patrol robot control method, which can be used for high-precision and real-time detection of floating fish in complex water surface background conditions. The overall research idea and process are shown in Figure 1 The main innovations are as follows:

[0095] (1) According to the characteristics of the complex interference background of the factory farming water surface, such as reflection, corrugation, low contrast and the like, a special data enhancement scheme is constructed, which includes water surface highlight simulation, corrugation disturbance, low contrast adjustment and the like, and multi-source acquisition video frame data is introduced in the training set, so that the original YOLOv12 model can learn and adapt to the visual feature distribution under the real farming environment in the training stage, thereby significantly improving the robustness of the model to the complex background.

[0096] (2) In order to adapt to the characteristics of the small target and sparse distribution of the floating head fish in the video monitoring picture, the multi-scale detection mechanism of the original YOLOv12 is used to increase the weight proportion of the high resolution feature layer in the training process, and the small target resampling is performed on the labeled data, so as to improve the detection recall rate of the original model to the small size fish body.

[0097] The method can maintain high-precision floating head detection under the complex background of factory farming, while taking into account the small target detection performance and real-time performance, and reduces false positives and omissions, providing a feasible and efficient solution for intelligent aquaculture management.

[0098] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A track-mounted intelligent fish monitoring robot control system, comprising a factory-style facility, characterized in that, M fish ponds are set up in the factory building, where M is a positive integer greater than or equal to 1; Inspection tracks are set up on the factory buildings for the movement of inspection robots, and the inspection robots are set on the tracks. It also includes industrial control computers and inspection platforms; After the inspection robot moves to the target fish pond, it takes pictures of the water surface and uploads the pictures to the edge industrial control computer. The industrial control computer identifies the number of dead fish and / or fish surfacing in the pictures and uploads the results to the digital twin management and control cloud platform.

2. The track-mounted intelligent fish monitoring robot control system according to claim 1, characterized in that, It also includes pushing an early warning message to the WeChat official account linked to the system when the number of dead fish exceeds a set number, providing timely reminders. The warning message is that the number of dead fish has exceeded the set number. Or / and set one or any combination of the daily inspection frequency, inspection photo points and charging time for the robot; Or / and M is 56, there are 7 fish ponds in a horizontal column and 8 fish ponds in a vertical column, for a total of 56 fish ponds; Or / and the inspection track is a circular distributed track; Or / and inspection robots are inspection robots equipped with surveillance cameras.

3. A control method for a track-based intelligent fish condition inspection robot, characterized in that, Includes the following steps: S1, The surveillance camera acquires and collects monitoring video of the fish pond surface; S2, transmit the video to the industrial control computer, which then cuts the frames to obtain images of the water surface; S3, Select images containing fish surfacing and / or dead fish from the images; use them as training, testing, and validation sets respectively, in proportion to the desired ratio; S4. Import the images from the training set into the YOLOv12 model to train the YOLOv12 model; import the images from the test set into the trained YOLOv12 model to test the trained YOLOv12 model; import the images from the validation set into the tested YOLOv12 model to validate the tested YOLOv12 model; finally, the target YOLOv12 model is obtained. S5 imports the target fishpond image captured by the inspection robot into the target YOLOv12 model and outputs the result.

4. The control method for the track-mounted intelligent fish condition inspection robot according to claim 3, characterized in that, In step S1, the fish ponds are fish ponds with different water quality and light conditions; Or / and the ratio of training set, test set and validation set in step S3 is 8:1:1; Or / and in step S3, the areas where fish are surfacing or / and dead are precisely marked using LabelImg software, instead of marking the entire fish. The reason for this is that... Fish surfacing behavior is determined by the feature of their heads emerging from the water. The fish head is relatively stable on the water surface and is less affected by water waves, which makes it easy for the model to accurately locate and detect it. Through this targeted labeling and scene-adaptive data collection method, combined with the model's multi-scale detection capabilities, it is possible to achieve efficient identification and real-time monitoring of fish surfacing and dead fish under complex water surface conditions.

5. The control method for the track-mounted intelligent fish condition inspection robot according to claim 3, characterized in that, In step S4, the YOLOv12 model includes a YOLOv12 backbone network, a YOLOv12 neck network, and a YOLOv12 detection head; The data output of the YOLOv12 backbone network is connected to the data input of the YOLOv12 neck network, and the data output of the YOLOv12 neck network is connected to the data input of the YOLOv12 detection head. The YOLOv12 backbone network is responsible for extracting multi-level feature information from the input image, providing basic feature representations for subsequent object detection; The YOLOv12 neck network is used to fuse and enhance multi-scale features extracted by the backbone network in order to detect targets of different sizes; The YOLOv12 detector head is used to convert the multi-scale fusion features output by the neck network into the final detection result.

6. The control method for the track-mounted intelligent fish condition inspection robot according to claim 3, characterized in that, The YOLOv12 backbone network includes an improved CSPDarknet structure. This improved CSPDarknet structure divides the feature flow into two parts through cross-stage partial connections. One part is directly passed to the next stage, while the other part is combined with the output of the next stage through skip connections. In this way, the gradient flow in the network is effectively separated, avoiding the gradient vanishing problem that may occur in traditional deep networks, while reducing the computational cost of each stage. In addition, the effective separation of the feature flow in CSPDarknet allows each part of the network to focus on feature learning at different levels. This not only improves the model's feature representation ability but also enhances the network's flexibility in handling complex tasks. Through this separation, CSPDarknet maintains high computational efficiency at each stage. CSPDarknet integrates the RepConv convolutional module and the SPPF module. RepConv is integrated into the convolutional layer to improve feature extraction performance, which can significantly improve the network's learning ability in deep networks. SPPF is used to enhance the network's receptive field, enabling it to capture richer spatial information while maintaining computational efficiency. The YOLOv12 neck network employs an improved PAN-FPN structure, effectively combining high-level semantic features with low-level fine-grained features. This preserves the detailed information of small targets while enhancing the semantic expression of large targets. Specifically, FPN is responsible for upsampling high-level features to lower levels and downsampling to enhance the semantic expression of low-level features, while PAN optimizes the information flow of low-level features, ensuring that detailed information can be efficiently transmitted to deeper network layers. After fusing these two paths, PAN-FPN combines high-level semantic information with low-level fine-grained information, effectively improving the accuracy of target detection tasks. PAN-FPN also introduces a C2f lightweight feature processing module, which improves information flow efficiency and reduces the risk of gradient vanishing through cross-layer connections and feature reuse. The YOLOv12 detection head includes the YOLO series' multi-scale detection strategy, which makes predictions on feature maps of different resolutions, while taking into account the detection needs of large, medium and small targets. The prediction branch at each scale outputs a tensor with a fixed number of channels through a convolutional layer, which includes the center coordinates, width and height of the bounding box, the target confidence, and the probability of each class. The YOLOv12 detection head uses an Anchor-Free mechanism during the decoding process to simplify anchor box matching and improve the detection accuracy of small targets and dense scenes; The YOLOv12 detection head works in conjunction with loss functions, including CIoU, DFL, classification cross-entropy, or any combination thereof. Through end-to-end optimization, a balance is achieved between detection performance and inference speed, ensuring real-time high-precision target detection.

7. The control method for the track-mounted intelligent fish condition inspection robot according to claim 3, characterized in that, Step S3 also includes loading YOLOv12 weights pre-trained on the COCO dataset, which allows the model to be fine-tuned based on existing general features, specifically adapted to the detection tasks of fish surfacing and dead fish. This transfer learning approach shortens training time while improving detection accuracy and generalization ability.

8. The control method for the track-mounted intelligent fish condition inspection robot according to claim 3, characterized in that, Step S3 also includes adopting a phased dynamic learning rate strategy to avoid the model getting stuck in local optima and ensure stable convergence; The optimizer is SGD, which can be flexibly adjusted according to the data size and convergence speed requirements. During the inference phase, the confidence threshold and NMS threshold are tuned; a lower threshold helps improve recall, while a higher threshold reduces false positives.

9. A computer system, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the control method for the track-type intelligent fish condition inspection robot as described in any one of claims 3 to 8 when executing the executable instructions.

10. A computer-readable storage medium, characterized in that, include: A memory on which computer programs are stored; A processor is configured to execute the program in the memory to implement the control method for the track-type intelligent fish monitoring robot as described in any one of claims 3 to 8.

Images