Intelligent monitoring method and system for grouper viral nervous necrosis phenotype

Abstract

The application discloses a grouper viral nervous necrosis symptom intelligent monitoring method and system, which comprises the following steps: collecting a grouper swimming video, and dividing the video into first and second standardized frame sequences after preprocessing; performing multi-target tracking on the first sequence, obtaining the unique ID and trajectory of each fish, and counting the total number of groupers; identifying sick fish based on the trajectory, and outputting the initial number of sick fish and the corresponding ID set; performing body surface feature detection on the second sequence after frame extraction, identifying sick fish with lesion characteristics, and outputting the lesion coordinates; and fusing and analyzing the recognition results of the two sequences, and outputting the final number of sick fish and the proportion. The application realizes automatic monitoring, and improves the detection efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of phenotypic patterns of diseases in aquaculture, specifically to an intelligent monitoring method and system for viral neuronecrosis phenotype in grouper. Background Technology

[0002] The grouper industry is severely impacted by viral neuronecrosis. Each year, this virus causes the loss of over 80% of total grouper fry, resulting in significant losses for the industry. The pathogen causing this disease is Red Spotted Grouper Neuronecrosis Virus (RGNNV), which primarily infects grouper fry. Typical symptoms after infection include abnormal swimming and blackening of the body. Currently, grouper disease detection mainly relies on manual observation, using the naked eye to examine the grouper's swimming behavior and body surface characteristics to determine if it is diseased. This traditional detection method has significant shortcomings: First, viral nerve necrosis mainly occurs in juvenile fish, and the small size of the fish makes it difficult to visually identify the disease phenotype. Second, it is highly subjective, with varying observation experience among different aquaculture personnel, leading to inconsistent accuracy in disease assessment and a high risk of missed or false detections. This is especially problematic when conducting antiviral breeding, as the inability to accurately identify disease phenotypes results in poor breeding precision. Third, it is inefficient, as manual observation cannot achieve real-time monitoring in large-scale aquaculture scenarios, and significant manpower is required for mortality statistics. Fourth, it suffers from delayed response, as manual observation makes it difficult to detect early-stage diseases, leading to disease spread and increasing the difficulty of prevention and control.

[0003] With the application of artificial intelligence technology in aquaculture, some image recognition-based technologies for precise identification of disease phenotypes have emerged. However, most existing technologies rely solely on surface features or movement characteristics for detection, resulting in limited accuracy. Furthermore, they fail to achieve precise counting and matching through multi-feature fusion, thus failing to meet the needs of precise identification and practical monitoring of grouper viral neuronecrosis phenotypes. Therefore, developing an automated, intelligent, and highly accurate method and system for precise identification of grouper viral neuronecrosis phenotypes has become an urgent technical problem to be solved.

[0004] Chinese invention application No. 202410095671.4 discloses "A Method and System for Early Warning of Fish Diseases," which includes the following steps: acquiring information on the fish's body surface, behavior, and water quality for feature training, and extracting feature vectors to reflect the fish's health status and the impact of diseases. YOLOv8 is used to classify the fish's body surface information, and YOLOv8+ByteTrack multi-target tracking technology is used to accurately detect and track fish targets, avoiding target loss and misidentification. LSTM time series prediction technology is used to predict the future state of fish and capture changes in fish behavior and abnormal situations. Fuzzy inference technology is used to reasonably set fuzzy sets and fuzzy rules, effectively perform fuzzy inference, and output the fish's disease early warning level. Summary of the Invention

[0005] To address the technical problems of existing grouper disease monitoring technologies, which rely on manual observation, are inefficient, subjective, and lack accuracy, and are difficult to implement in real-time monitoring in large-scale grouper farming, this invention provides an intelligent monitoring method and system for the phenotype of viral neuronecrosis in grouper. The technical solution adopted by this invention is as follows:

[0006] The first aspect of this invention provides a method for intelligent monitoring of the phenotype of viral neuronecrosis in grouper, the method comprising: Collect video data of grouper swimming; Preprocessing of grouper swimming video data yields a first-channel standardized frame sequence and a second-channel standardized frame sequence. Multi-target tracking is performed on each frame of the first standardized frame sequence to obtain the unique identifier ID and continuous coordinate trajectory of each grouper, and the total number of groupers in the current video sequence is counted based on the unique identifier ID. Based on the continuous coordinate trajectory, kinematic features related to viral neuronecrosis are extracted, and diseased fish are identified according to the kinematic features. The first diseased fish identification result is output, which includes the initial number of diseased fish and their corresponding ID set. The image frames in the second standardized frame sequence are extracted, and the surface features of the extracted images are detected to identify the surface lesion features related to viral nerve necrosis. The second diseased fish identification result is output, which includes the coordinates of the diseased grouper. The results of the first diseased fish identification and the second diseased fish identification are fused and analyzed to output the final number and percentage of diseased fish.

[0007] As a preferred embodiment, the preprocessing includes: Scale the grouper swimming video to a uniform resolution; The grouper swimming video was decoded frame by frame at the original frame rate to obtain a standardized frame sequence. The standardized frame sequence is divided into two outputs to obtain a first standardized frame sequence and a second standardized frame sequence; wherein, the first standardized frame sequence is a high frame rate stream, and the second standardized frame sequence is a low frame rate stream after downsampling.

[0008] As a preferred embodiment, the multi-target tracking employs the FairMOT algorithm, which uses a joint learning framework of detection and re-identification. The detection branch outputs target bounding boxes, the embedding branch learns appearance features for target re-identification, and Kalman filtering and the Hungarian algorithm are used for data association to achieve consistent ID allocation across frames.

[0009] As a preferred embodiment, the method for extracting kinematic features related to viral neuronecrosis based on the continuous coordinate trajectory, identifying diseased fish according to the kinematic features, and outputting a first diseased fish identification result includes: Receive the unique identifier (ID) and continuous coordinate trajectory of each grouper. Let the coordinate sequence of a certain ID within N consecutive frames be:

[0010] in, The coordinates of the bounding box center are Width and height, i =1,2,…, N The input tensor is ; Multi-branch feature extraction was performed on the continuous coordinate trajectory of each grouper to obtain local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features; The local motion pattern features, the long-range temporal dependency features, and the spatiotemporal hybrid features are concatenated to obtain the fused features. ; The fused features are input into an XGBoost classifier, which consists of K decision trees, with the k-th tree outputting a score. Ultimately, the confidence level of the diseased fish for:

[0011] in For the Sigmoid function; Set confidence threshold ,like Then the grouper corresponding to that ID is determined to be a diseased fish.

[0012] As a preferred embodiment, the method for extracting multi-branch features from the continuous coordinate trajectory of each grouper to obtain local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features includes: The continuous coordinate trajectory of each grouper is used to extract features through CNN sub-network branches, Bi-LSTM sub-network branches, and CNN-LSTM sub-network branches to obtain the corresponding local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features.

[0013] As a preferred embodiment, the surface feature detection employs the D-FINE target detection model.

[0014] As a preferred embodiment, the fusion analysis method includes: The coordinates of the diseased grouper detected in the second diseased fish identification result are compared with the coordinates of the grouper obtained by multi-target tracking in the same frame by performing an intersection-union calculation. Based on the crossover ratio calculation result, the second diseased fish identification result is assigned to the corresponding tracking ID; Based on the first and second diseased fish identification results corresponding to each tracking ID, the final total number of diseased fish is determined according to the preset fusion decision logic.

[0015] As a preferred embodiment, the fusion decision logic includes: If the first diseased fish corresponding to an ID is identified as normal while the second diseased fish is identified as lesioned, then it is considered a supplementary diseased fish; if the kinematics determines that it is lesioned, then regardless of the body surface determination result, it is not considered a supplementary diseased fish. Finally, the number of diseased fish added is counted, and the sum of the initial number of diseased fish and the number of added diseased fish is taken as the final total number of diseased fish. Based on the final total number of diseased fish and the total number of grouper, the percentage of diseased fish is obtained and the result is output.

[0016] The second aspect of this invention provides an intelligent monitoring system for grouper diseases, the system comprising a video acquisition module, a preprocessing module, a multi-target tracking module, a kinematic recognition module, a body surface feature recognition module, and a multi-mode fusion module; The video acquisition module is used to collect video data of grouper swimming; The preprocessing module is used to preprocess the grouper swimming video data to obtain a first normalized frame sequence and a second normalized frame sequence. The multi-target tracking module is used to perform multi-target tracking on each frame of the first standardized frame sequence, obtain the unique identifier ID and continuous coordinate trajectory of each grouper, and count the total number of groupers in the current video sequence based on the unique identifier ID. The kinematic recognition module is used to extract kinematic features related to viral neuronecrosis based on the continuous coordinate trajectory, and to identify diseased fish according to the kinematic features, and output a first diseased fish identification result; the first diseased fish identification result includes the initial number of diseased fish and their corresponding ID set; The body surface feature recognition module is used to perform frame extraction processing on the image frames in the second standardized frame sequence, perform body surface feature detection on the extracted images, identify body surface lesion features related to viral nerve necrosis, and output the second diseased fish identification result; the second diseased fish identification result includes the coordinates of the diseased grouper; The polymorphic fusion module is used to fuse and analyze the first diseased fish identification result and the second diseased fish identification result, and output the final result of the number and proportion of diseased fish.

[0017] A third aspect of the present invention provides a computer device, including a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the aforementioned intelligent monitoring method for the phenotype of grouper viral neuronecrosis.

[0018] Compared with the prior art, the beneficial effects of this invention are: This invention employs a three-stage "tracking-detection-fusion" technical architecture, replacing traditional manual visual observation. It enables 24 / 7 continuous monitoring without significant manual intervention, greatly improving monitoring efficiency. This approach is suitable for large-scale disease-resistant breeding and production of fossilized grouper, solving the technical problems of low efficiency and difficulty in large-scale monitoring associated with manual observation. The invention performs parallel detection from two dimensions by fusing kinematic features and surface lesion features. It combines a self-developed multi-branch network with XGBoost fusion classification to accurately extract complex behavioral features. Furthermore, it utilizes IoU matching and fusion decision logic to achieve accurate ID association and deduplication counting, comprehensively improving the accuracy of anomaly identification and diseased fish counting, overcoming the problems of strong subjectivity and insufficient accuracy of single-feature identification in manual observation. The invention employs a data splitting strategy: a high-frame-rate stream ensures continuous motion tracking, while a downsampling stream reduces the computational load of surface detection, enabling multi-module parallel inference. This significantly reduces computational consumption while maintaining recognition accuracy, achieving real-time monitoring and early disease warning, solving the problem of delayed response in manual observation. Attached Figure Description

[0019] Figure 1 This embodiment provides a flowchart of a method for intelligent monitoring of the phenotype of viral neuronecrosis in grouper. Figure 2 This is a schematic diagram of the kinematic recognition module structure provided in this embodiment; Figure 3The diagram shows the application effect of the intelligent monitoring method for the phenotype of viral neuronecrosis in grouper provided in this embodiment. Detailed Implementation

[0020] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. It should be understood that the described embodiments are merely some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0021] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0022] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. The invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Example 1 Please refer to Figure 1 This embodiment provides a method for intelligent monitoring of the phenotype of viral neuronecrosis in grouper, the method comprising: S1: Collect video data of grouper swimming; S2: Preprocess the grouper swimming video data to obtain the first and second normalized frame sequences; In one specific embodiment, the preprocessing includes: Scale the grouper swimming video to a uniform resolution; Specifically, the input video is scaled to a uniform resolution (e.g., 1280×960) to eliminate resolution differences caused by different acquisition devices; The grouper swimming video was decoded frame by frame at the original frame rate to obtain a standardized frame sequence. The standardized frame sequence is divided into two outputs to obtain a first standardized frame sequence and a second standardized frame sequence; wherein, the first standardized frame sequence is a high frame rate stream, and the second standardized frame sequence is a low frame rate stream after downsampling.

[0026] S3: Perform multi-target tracking on each frame of the first standardized frame sequence, obtain the unique identifier ID and continuous coordinate trajectory of each grouper, and count the total number of groupers in the current video sequence based on the unique identifier ID. In one specific embodiment, the multi-target tracking uses the FairMOT algorithm, which employs a joint learning framework of detection-re-identification. The detection branch outputs target bounding boxes, the embedding branch learns appearance features for target re-identification, and Kalman filtering and the Hungarian algorithm are used for data association to achieve cross-frame ID consistency allocation.

[0027] S4: Based on the continuous coordinate trajectory, extract the kinematic features related to viral neuronecrosis, identify diseased fish according to the kinematic features, and output the first diseased fish identification result, which includes the initial number of diseased fish and their corresponding ID set; Specifically, the kinematic features associated with viral neuronecrosis include one or more of the following: continuous rotational swimming, abnormal trajectory curvature, prolonged stillness, rotational swimming along the body's long axis, or aimless swimming.

[0028] In one specific embodiment, the method for extracting kinematic features related to viral neuronecrosis based on the continuous coordinate trajectory, identifying diseased fish according to the kinematic features, and outputting a first diseased fish identification result includes: Receive the unique identifier (ID) and continuous coordinate trajectory of each grouper. Let the coordinate sequence of a certain ID within N consecutive frames be:

[0029] in, The coordinates of the bounding box center are Width and height,i =1,2,…, N The input tensor is ; Multi-branch feature extraction was performed on the continuous coordinate trajectory of each grouper to obtain local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features; The local motion pattern features, the long-range temporal dependency features, and the spatiotemporal hybrid features are concatenated to obtain the fused features. ; The fused features are input into an XGBoost classifier, which consists of K decision trees, with the k-th tree outputting a score. Ultimately, the confidence level of the diseased fish for:

[0030] in For the Sigmoid function; Set confidence threshold ,like Then the grouper corresponding to that ID is determined to be a diseased fish.

[0031] In a specific embodiment, the method for extracting multi-branch features from the continuous coordinate trajectory of each grouper to obtain local motion pattern features, long-distance temporal dependency features, and spatiotemporal hybrid features includes: The continuous coordinate trajectory of each grouper is used to extract features through CNN sub-network branches, Bi-LSTM sub-network branches, and CNN-LSTM sub-network branches to obtain the corresponding local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features.

[0032] Specifically, the CNN sub-network branch employs stacked one-dimensional convolutional layers, with the convolutional kernels sliding along the time dimension to capture local dynamic features such as instantaneous velocity changes, acceleration, and trajectory curvature. The Bi-LSTM sub-network branch uses a bidirectional long short-term memory network; the forward LSTM captures motion trends from the past to the future, while the backward LSTM captures motion patterns from the future to the past, used to identify behavioral features requiring long-term observation, such as "long-term stillness" and "periodic abnormal swimming." The CNN-LSTM sub-network branch first extracts local spatial features through one-dimensional convolutional layers, then inputs the convolutional output sequence into the LSTM layer, fusing spatial locality and temporal memory capabilities to identify complex composite behaviors such as "rotational swimming" and "spiral trajectories."

[0033] S5: Perform frame extraction processing on the image frames in the second standardized frame sequence, perform body surface feature detection on the extracted images, identify body surface lesion features related to viral nerve necrosis, and output the second diseased fish identification result, which includes the coordinates of the diseased grouper. Specifically, the skin lesion features associated with viral neuronecrosis include one or more of the following: darkened skin color, pale skin color, skin ulceration, white spots, bulging eyes, or abdominal swelling.

[0034] In one specific embodiment, the body surface feature detection employs the D-FINE target detection model.

[0035] Specifically, the D-FINE model employs a dynamic feature refinement mechanism, utilizing a weighted bidirectional feature pyramid network (BiFPN) and a dynamic label allocation strategy to achieve high-precision target detection. The detection output includes bounding box coordinates. Category label (normal / pathological), confidence score.

[0036] Specifically, the frame extraction method includes: Fixed-interval frame extraction: 1 frame is extracted every M frames (e.g., 1 frame is extracted every 3 frames). Adaptive frame extraction: dynamically adjusts the sampling rate based on the abnormal confidence output by the kinematic recognition module to improve the sampling rate in areas suspected of being diseased fish.

[0037] S6: Combine and analyze the first diseased fish identification result with the second diseased fish identification result, and output the final result of the number and percentage of diseased fish; In one specific embodiment, the fusion analysis method includes: The coordinates of the diseased grouper detected in the second diseased fish identification result are compared with the coordinates of the grouper obtained by multi-target tracking in the same frame by performing an intersection-union calculation. Specifically, for each diseased grouper identified by body surface feature recognition, the coordinate frame is... Calculate its tracking bounding boxes and all tracking boxes output by multi-target tracking in the same frame. Intersection over Union (IoU):

[0038] Set matching threshold (e.g., 0.99), if The lesion feature will then be assigned to the corresponding tracking ID.

[0039] Based on the crossover ratio calculation result, the second diseased fish identification result is assigned to the corresponding tracking ID; Based on the first and second diseased fish identification results corresponding to each tracking ID, the final total number of diseased fish is determined according to the preset fusion decision logic.

[0040] In one specific embodiment, the fusion decision logic includes: If the first diseased fish corresponding to an ID is identified as normal while the second diseased fish is identified as lesioned, then it is considered a supplementary diseased fish; if the kinematics determines that it is lesioned, then regardless of the body surface determination result, it is not considered a supplementary diseased fish. Finally, the number of diseased fish added is counted, and the sum of the initial number of diseased fish and the number of diseased fish added is taken as the final total number of diseased fish. Based on the final total number of diseased fish and the total number of grouper, the percentage of diseased fish is obtained and the result is output. Specifically, for each tracking ID, there are four possible state combinations, as shown in Table 1: Table 1. Four state combinations for each tracking ID

[0041] The specific handling logic for the corresponding situation is as follows: Let n be the number of IDs appearing in the current sequence (100 frames), m be the number of diseased fish identified by kinematics, and m be the total number of diseased fish for now.

[0042] The grouper coordinates of the diseased grouper are extracted and identified as the category of the surface feature. The coordinates of the diseased grouper are compared one by one with the grouper coordinates output by the multi-target tracking in the same frame. If the intersection-union ratio (IOU) of the diseased grouper coordinates and any grouper coordinates output by the multi-target tracking exceeds 0.99, the grouper ID corresponding to the coordinates is extracted to complete the matching.

[0043] For case 1, no action is taken.

[0044] For scenario 2, update the total number of diseased fish to m+1.

[0045] No action is taken for scenario 3.

[0046] No action is taken for scenario 4.

[0047] After processing all IDs in the current sequence according to the above steps, the number and percentage of diseased fish are calculated and output based on the updated total number of diseased fish and the total number of grouper identified in the video.

[0048] Example 2 This embodiment provides an intelligent monitoring system for grouper diseases. The system includes a video acquisition module, a preprocessing module, a multi-target tracking module, a kinematic recognition module, a body surface feature recognition module, and a multi-mode fusion module. The video acquisition module is used to collect video data of grouper swimming; The preprocessing module is used to preprocess the grouper swimming video data to obtain a first normalized frame sequence and a second normalized frame sequence. The multi-target tracking module is used to perform multi-target tracking on each frame of the first standardized frame sequence, obtain the unique identifier ID and continuous coordinate trajectory of each grouper, and count the total number of groupers in the current video sequence based on the unique identifier ID. The kinematic recognition module is used to extract kinematic features related to viral neuronecrosis based on the continuous coordinate trajectory, and to identify diseased fish according to the kinematic features, and output a first diseased fish identification result; the first diseased fish identification result includes the initial number of diseased fish and their corresponding ID set; For details, please refer to Figure 2 The kinematic recognition module includes parallel CNN sub-network branches, Bi-LSTM sub-network branches, CNN-LSTM sub-network branches, and an XGBoost fusion network; the inputs of the CNN sub-network branches, Bi-LSTM sub-network branches, and CNN-LSTM sub-network branches are connected to the multi-target tracking module, and the outputs are connected to the input of the XGBoost fusion network; the XGBoost fusion network includes a fusion module and an XGBoost classifier; The fusion module is used to concatenate the local motion pattern features, the long-range temporal dependency features, and the spatiotemporal hybrid features to obtain fused features. ; The XGBoost classifier is used to receive fused features. Output the final confidence score of the diseased fish. .

[0049] The body surface feature recognition module is used to perform frame extraction processing on the image frames in the second standardized frame sequence, perform body surface feature detection on the extracted images, identify body surface lesion features related to viral nerve necrosis, and output the second diseased fish identification result; the second diseased fish identification result includes the coordinates of the diseased grouper; The polymorphic fusion module is used to fuse and analyze the first diseased fish identification result and the second diseased fish identification result, and output the final result of the number and proportion of diseased fish.

[0050] Specifically, in this embodiment, the grouper farming scenario is an indoor factory-style aquaculture pond, in which grouper with a body length of 2-5cm are stocked. A high-definition camera (supporting 4:3 aspect ratio shooting) is installed above the pond as a video acquisition module. The camera angle is adjusted to ensure that the swimming status of all grouper in the pond can be captured completely. The camera is turned on to collect video of the grouper swimming, and the video frame rate is set to 30 frames per second. Feeding time is avoided to reduce interference with feeding behavior. Each pond is continuously recorded for 20 minutes, and a total of 60 video clips are collected, of which 40 clips are used for training and 20 clips are used for testing.

[0051] Specifically, annotation methods for multi-object tracking datasets include: Label the minimum bounding rectangle for all groupers in each frame of the video. Labeling specifications: Bounding box rules: The bounding box must fit snugly against the fish's outline and must not contain shadows or reflections; ID allocation rule: Assign a unique integer ID (starting from 1) to each fish, and keep the ID consistent across frames; Occlusion handling: When the fish body is occluded by more than 50%, it is marked as "occluded" but retains its original ID; Entry and exit processing: If a fish swims out of the screen and then re-enters, and the time interval is less than 1 second and the appearance characteristics match, the original ID is restored.

[0052] After annotation, export as a MOT Challenge file (*.txt), with each line formatted as follows:<frame_id> ,<track_id> , <x> , <y> , <w> , <h> , <confidence> , <class> , <visibility>Where coordinates The coordinates of the top left corner of the bounding box. The width and height are represented by the value, and visibility indicates the visibility (0-1).

[0053] Specifically, the annotation methods for body surface feature detection datasets include: Based on the same batch of videos, one frame out of every three frames is extracted as a keyframe for annotation. Category definition: Category 0 (Normal): Normal body color (brown or yellowish-brown), no ulcers, white spots, or swelling on the body surface; Category 1 (lesions): The body color is black or white, and visible lesions such as ulceration, white spots, bulging eyeballs, and abdominal swelling appear on the body surface.

[0054] The annotations are formatted using the COCO dataset and exported as JSON files, containing image information, category information, and annotation information (bounding box coordinates and category ID).

[0055] Specifically, the methods for constructing kinematic recognition datasets include: The kinematic recognition dataset is generated semi-automatically, eliminating the need for manual frame-by-frame annotation. (1) Run the trained FairMOT model to process the original video and generate tracking results; (2) Manually review the tracking results, correct ID switching errors, and ensure the continuity of the trajectory; (4) Extract the coordinate sequence of N=100 consecutive frames by ID as a sample; (4) Manually observe the movement behavior of this ID within 100 frames and label it according to the following standards: Abnormal label (1): Any of the following behaviors occur: continuously spinning on the water surface, hitting the pool wall more than 3 times in a row, remaining still for more than 10 seconds, rotating and swimming along the body's long axis, swimming randomly without direction (with disordered and intersecting trajectories). Normal label (0): Swims smoothly with a clear direction, occasionally turns, and does not exhibit the above-mentioned abnormal behaviors.

[0056] Specifically, the training method for the multi-target tracking module includes: Based on the PaddleDetection framework, the FairMOT algorithm (configuration file fairmot_dla34_30e_1088x608.yml) was selected for training, with the input size modified to 1280x960, and the other hyperparameters using the values ​​from the original configuration file. After training, the model was exported in ONNX format.

[0057] Specifically, the training method for the body surface feature recognition module includes: Based on the official D-FINE open-source framework configuration, the input size was modified to 1280x960, and the other hyperparameters used the values ​​in the original configuration file. After training, the model was exported in ONNX format.

[0058] Specifically, the training method for the kinematic recognition module includes: The detailed architecture of the kinematics recognition module is shown in the appendix. Figure 1 The construction process is as follows: The XGBoost classifier in the kinematic recognition module is trained.

[0059] The features output from the three sub-network branches are combined and used as the input to XGBoost. The XGBoost hyperparameters are set as follows: Number of trees (n_estimators): 200 Maximum depth (max_depth): 6 Learning rate: 0.1 Subsampling rate: 0.8 Column sampling rate (colsample_bytree): 0.8 Regularization parameters (reg_alpha): 0.1, (reg_lambda): 1.0 Training process: Two-stage training is employed: Phase 1 (End-to-end training): Fixed XGBoost, only three sub-network branches are trained, cross-entropy loss is optimized, and training lasts for 50 epochs; Phase 2 (Feature Extraction + Classifier Training): Freeze sub-network branches, extract training set features, and train the XGBoost classifier.

[0060] Please refer to Figure 3 This invention enables automated and intelligent identification of grouper disease (viral neuronecrosis), significantly improving the efficiency and accuracy of phenotypic identification of the disease, reducing the error and cost of manual identification, and is adaptable to large-scale grouper disease-resistant breeding and production aquaculture scenarios, providing real-time and reliable support for accurate phenotypic identification of grouper viral neuronecrosis.

[0061] Example 3 This embodiment provides a computer device, including a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, it implements the steps of the intelligent monitoring method for the phenotype of grouper viral neuronecrosis described in Embodiment 1.

[0062] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.< / visibility> < / class> < / confidence> < / h> < / w> < / y> < / x>

Claims

1. A method for intelligent monitoring of the phenotype of viral neuronecrosis in grouper, characterized in that, The method includes: Collect video data of grouper swimming; Preprocessing of grouper swimming video data yields a first-channel standardized frame sequence and a second-channel standardized frame sequence. Multi-target tracking is performed on each frame of the first standardized frame sequence to obtain the unique identifier ID and continuous coordinate trajectory of each grouper, and the total number of groupers in the current video sequence is counted based on the unique identifier ID. Based on the continuous coordinate trajectory, kinematic features related to viral neuronecrosis are extracted, and diseased fish are identified according to the kinematic features. The first diseased fish identification result is output, which includes the initial number of diseased fish and their corresponding ID set. The image frames in the second standardized frame sequence are extracted, and the surface features of the extracted images are detected to identify the surface lesion features related to viral nerve necrosis. The second diseased fish identification result is output, which includes the coordinates of the diseased grouper. The results of the first diseased fish identification and the second diseased fish identification are fused and analyzed to output the final number and percentage of diseased fish. The preprocessing includes: Scale the grouper swimming video to a uniform resolution; The grouper swimming video was decoded frame by frame at the original frame rate to obtain a standardized frame sequence. The standardized frame sequence is divided into two outputs to obtain a first standardized frame sequence and a second standardized frame sequence; wherein, the first standardized frame sequence is a high frame rate stream, and the second standardized frame sequence is a low frame rate stream after downsampling; The fusion analysis method includes: The coordinates of the diseased grouper detected in the second diseased fish identification result are compared with the coordinates of the grouper obtained by multi-target tracking in the same frame by performing an intersection-union calculation. Based on the crossover ratio calculation result, the second diseased fish identification result is assigned to the corresponding tracking ID; Based on the first and second diseased fish identification results corresponding to each tracking ID, the final total number of diseased fish is determined according to the preset fusion decision logic; The fusion decision logic includes: If the first diseased fish corresponding to an ID is identified as normal while the second diseased fish is identified as lesioned, then it is considered a supplementary diseased fish; if the kinematics determines that it is lesioned, then regardless of the body surface determination result, it is not considered a supplementary diseased fish. Finally, the number of diseased fish added is counted, and the sum of the initial number of diseased fish and the number of added diseased fish is taken as the final total number of diseased fish. Based on the final total number of diseased fish and the total number of grouper, the percentage of diseased fish is obtained and the result is output.

2. The intelligent monitoring method for the phenotype of viral neuronecrosis in grouper according to claim 1, characterized in that, The multi-target tracking uses the FairMOT algorithm, which employs a joint learning framework of detection and re-identification. The detection branch outputs target bounding boxes, and the embedding branch learns appearance features for target re-identification. Kalman filtering and the Hungarian algorithm are used for data association to achieve consistent ID allocation across frames.

3. The intelligent monitoring method for the phenotype of viral neuronecrosis in grouper according to claim 1, characterized in that, The method for extracting kinematic features related to viral neuronecrosis based on the continuous coordinate trajectory, identifying diseased fish according to the kinematic features, and outputting the first diseased fish identification result includes: Receive the unique identifier (ID) and continuous coordinate trajectory of each grouper. Let the coordinate sequence of a certain ID within N consecutive frames be: in, The coordinates of the bounding box center are Width and height, i =1,2,…, N The input tensor is ; Multi-branch feature extraction was performed on the continuous coordinate trajectory of each grouper to obtain local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features; The local motion pattern features, the long-range temporal dependency features, and the spatiotemporal hybrid features are concatenated to obtain the fused features. ; The fused features are input into an XGBoost classifier, which consists of K decision trees, with the k-th tree outputting a score. Ultimately, the confidence level of the diseased fish for: in For the Sigmoid function; Set confidence threshold ,like Then the grouper corresponding to that ID is determined to be a diseased fish.

4. The intelligent monitoring method for the phenotype of viral neuronecrosis in grouper according to claim 3, characterized in that, The method for extracting multi-branch features from the continuous coordinate trajectory of each grouper to obtain local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features includes: The continuous coordinate trajectory of each grouper is used to extract features through CNN sub-network branches, Bi-LSTM sub-network branches, and CNN-LSTM sub-network branches to obtain the corresponding local motion pattern features, long-distance temporal dependence features, and spatiotemporal hybrid features.

5. The intelligent monitoring method for the phenotype of viral neuronecrosis in grouper according to claim 1, characterized in that, The surface feature detection uses the D-FINE target detection model.

6. A grouper disease intelligent monitoring system, used to execute the intelligent monitoring method for the phenotype of grouper viral neuronecrosis as described in claim 1, characterized in that, The system includes a video acquisition module, a preprocessing module, a multi-target tracking module, a kinematic recognition module, a body surface feature recognition module, and a multi-mode fusion module; The video acquisition module is used to collect video data of grouper swimming; The preprocessing module is used to preprocess the grouper swimming video data to obtain a first normalized frame sequence and a second normalized frame sequence. The multi-target tracking module is used to perform multi-target tracking on each frame of the first standardized frame sequence, obtain the unique identifier ID and continuous coordinate trajectory of each grouper, and count the total number of groupers in the current video sequence based on the unique identifier ID. The kinematic recognition module is used to extract kinematic features related to viral neuronecrosis based on the continuous coordinate trajectory, and to identify diseased fish according to the kinematic features, and output a first diseased fish identification result; the first diseased fish identification result includes the initial number of diseased fish and their corresponding ID set; The body surface feature recognition module is used to perform frame extraction processing on the image frames in the second standardized frame sequence, perform body surface feature detection on the extracted images, identify body surface lesion features related to viral nerve necrosis, and output the second diseased fish identification result; the second diseased fish identification result includes the coordinates of the diseased grouper; The polymorphic fusion module is used to fuse and analyze the first diseased fish identification result and the second diseased fish identification result, and output the final result of the number and proportion of diseased fish.

7. A computer device, characterized in that: The method includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the intelligent monitoring method for the phenotype of grouper viral neuronecrosis as described in any one of claims 1 to 5.

Images