Lightweight fish body surface disease detection method and system based on improved RT-DETR

Abstract

The embodiment of the application relates to a fish body surface disease intelligent detection method and system based on a lightweight Transform architecture, the system integrates deep learning and computer vision technology, and constructs a high-precision, low-computing-complexity real-time disease recognition solution suitable for an aquaculture scene. The system adopts an optimized RT-DETR model, enhances the feature extraction capability for irregular lesions through a lightweight CSPDarknet backbone network and an innovative C2f-DSAN hybrid architecture, and introduces a PolaFormer polarity perception attention mechanism to improve the sensitivity of the model to subtle feature differences. The application is suitable for factory and intensive aquaculture scenes, can effectively deal with the risk of rapid disease spread in high-density breeding environments, provides real-time and reliable intelligent monitoring means for fish health management, and has good deployment adaptability and economic benefits.

Description

Technical Field

[0001] This invention belongs to the field of marine information technology, and specifically relates to a lightweight model and system for detecting diseases on the body surface of fish. Background Technology

[0002] With the intensive and large-scale development of global aquaculture, it has become an important pillar for ensuring food security and promoting economic growth. However, high-density farming has also exacerbated the risk of fish disease transmission. Surface diseases, as common and frequently occurring conditions, not only lead to huge economic losses but may also trigger ecological security issues such as antibiotic overuse. Therefore, real-time and accurate monitoring and early warning of the health status of fish surfaces are crucial for achieving the sustainable development of aquaculture.

[0003] Currently, the monitoring of fish surface diseases mainly relies on traditional manual observation and methods based on traditional machine learning. Manual observation depends on the experience of fish farmers and has inherent drawbacks such as high subjectivity, low efficiency, and inability to achieve 24 / 7 monitoring, making it difficult to meet the needs of large-scale fish farms. Although automated detection technology based on computer vision offers a solution, existing methods still face significant challenges:

[0004] Insufficient model performance: Typical detection models based on convolutional neural networks (CNN), such as YOLO and Faster R-CNN, have poor adaptability to irregularly shaped lesions and are not sensitive to subtle changes in features, resulting in high rates of missed detection and false detection.

[0005] High computational complexity: Some high-performance models, such as the DETR series based on Transformer, have a huge number of parameters and computational load, making it difficult to deploy efficiently on edge devices with limited computing resources and failing to meet the urgent need for real-time monitoring.

[0006] Incomplete feature interaction: The optimization strategies used in existing lightweight attempts, such as linear attention mechanisms, often sacrifice the model's ability to fully capture feature information in order to reduce computational complexity. In particular, they ignore negative feature interactions, which leads to a decrease in the model's discrimination ability and affects detection accuracy.

[0007] In summary, existing technologies struggle to achieve a good balance between detection accuracy, lightweight models, and real-time deployment. Therefore, there is an urgent need in this field for a lightweight intelligent detection solution for fish surface diseases that can balance high accuracy and high efficiency, enabling early detection, early warning, and early treatment of diseases, thereby promoting the intelligent upgrading of aquaculture. Summary of the Invention

[0008] 1. The purpose of this invention is to address the shortcomings of existing technologies by providing a lightweight model and system for detecting diseases on the surface of fish, thereby solving the problems mentioned in the background.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a lightweight method and system for detecting fish surface diseases based on improved RT-DETR, specifically including the following steps:

[0010] S1: Use underwater camera equipment to collect image data of fish body surface, covering a variety of common fish body surface diseases, including but not limited to ulcers, gill rot, fin damage, eye diseases, etc.

[0011] S2: The collected data is stored and managed through a high-performance data processing platform to ensure data integrity and accessibility;

[0012] S3: Construct an end-to-end object detection model PDFD-DETR based on the Transformer architecture. Integrate the lightweight variable convolution module DSANBlock and the PolaFormer polarity-aware attention mechanism into the RT-DETR model to enhance the model's adaptability to irregular lesion areas on the fish body surface and its ability to capture lesion details.

[0013] S4: A lesion visualization and analysis tool that inputs images of fish to be detected into the system and outputs location and classification information of disease areas, enhancing the interpretability of results, supporting real-time monitoring and decision-making applications, and enhancing interactivity and comprehensibility.

[0014] 2. As a preferred embodiment of the present invention: the underwater fish health status data acquisition system in S1 consists of the following modules:

[0015] A: Underwater high-definition camera equipment;

[0016] B: Underwater auxiliary lighting system;

[0017] C: Waterproof housing and pressure protection;

[0018] D: Underwater support and stabilization device;

[0019] E: Control and monitoring system;

[0020] 3. As a preferred embodiment of the present invention: the underwater high-definition camera device in A comprises the following parts:

[0021] A1 High-Definition Imaging Sensor: Employs a CMOS or CCD sensor, supporting a resolution of at least 4K (3840×2160 pixels) and a frame rate of ≥30fps, ensuring the capture of fish swimming dynamics and subtle lesions on their body surface;

[0022] A2 optical zoom lens group: includes autofocus (AF) module, optical image stabilization (OIS) mechanism and variable aperture (F1.8-F16), focal length range covers 20-200mm, supports macro shooting;

[0023] A3 Image Signal Processor: Integrated ISP chip for real-time color correction, noise reduction, and HDR synthesis;

[0024] A4 Data Interface Module: Provides Gigabit Ethernet, HDMI 2.0 and USB 3.0 interfaces, and supports raw data (RAW) format output;

[0025] A5 Environmental Adaptation Components: Includes automatic defogger heating element, hydrophobic coated lens and anti-biofouling electrode;

[0026] 4. As a preferred embodiment of the present invention: the underwater auxiliary lighting system in B comprises the following parts:

[0027] B1 Multispectral LED Array: Includes white, blue and ultraviolet LEDs, with adjustable power from 0-100W;

[0028] B2 Optical Diffusing System: Employs a combination of Fresnel lenses and diffusers to achieve 120° uniform illumination and eliminate hotspots;

[0029] B3 Synchronization Trigger Module: Achieves microsecond-level synchronization with the camera shutter through opto-isolation circuit, supporting front curtain / rear curtain synchronization modes;

[0030] B4 Intelligent Dimming System: Includes a light intensity sensor feedback loop, which automatically adjusts the output power according to the shooting distance;

[0031] B5 Heat Dissipation and Protection Components: Integrated copper-based heat pipe heat dissipation system and corrosion-resistant sapphire protective cover;

[0032] 5. As a preferred embodiment of the present invention: the waterproof outer shell and pressure protection in C consist of the following parts:

[0033] C1 Main Sealing System: Adopts a double O-ring radial sealing structure, the material is fluororubber (FKM), and the pressure resistance depth is ≥100 meters;

[0034] C2 pressure balancing device: includes silicone oil compensator and pressure relief valve, adaptable to pressure changes of 0.1-1.5MPa;

[0035] C3 structural protective frame: It adopts a combination design of 316L stainless steel main frame and polycarbonate observation window;

[0036] C4 Interface Waterproof Module: All electrical interfaces use wet-plug connectors. C5 Corrosion Protection System: Includes sacrificial anode blocks and anti-biodegradation coating.

[0037] 6. As a preferred embodiment of the present invention: the underwater support and stabilization device in D consists of the following parts:

[0038] D1 multi-degree-of-freedom gimbal: provides three-axis electric adjustment for pitch (±90°), roll (±45°) and translation;

[0039] D2 Adaptive Counterweight System: Includes adjustable counterweights and buoyancy materials to achieve neutral buoyancy control;

[0040] D3 Rapid Deployment Mechanism: Modular plug-in structure, supporting shipboard lifting, shore-based fixing, and robot mounting;

[0041] D4 shock absorption and flow-resistant components: integrates a fluid dynamics guide cover and a silicone oil damping shock absorber;

[0042] D5 position feedback module: includes tilt sensor, electronic compass and depth encoder;

[0043] 7. As a preferred embodiment of the present invention: the control and monitoring system in E comprises the following parts:

[0044] E1 main control computing unit: equipped with an ARM Cortex-A72 processor and FPGA coprocessor, running a Linux system;

[0045] E2 Remote Communication Module: Supports multi-mode transmission of 4G / 5G cellular networks, underwater acoustic communication and satellite communication;

[0046] E3 sensor integrated interface: Provides RS-485, CAN bus and I2C interfaces to connect water temperature, turbidity and depth sensors;

[0047] E4 Power Management System: Intelligent power distribution module, supports POE++ power supply and lithium battery pack, with a battery life of ≥72 hours;

[0048] E5 Status Monitoring Terminal: Real-time display of parameters such as device attitude, power status, and water leakage detection;

[0049] 8. As a preferred embodiment of the present invention: the high-performance data processing platform in S2 consists of the following modules:

[0050] V1: Data receiving and access layer;

[0051] V2: Distributed storage system;

[0052] V3: Data Management Module;

[0053] V4: Security and Access Control Module;

[0054] V5: Monitoring and Maintenance Subsystem;

[0055] V6: Integration and API Layer;

[0056] 9. As a preferred embodiment of the present invention: the data receiving and access layer in V1 consists of the following parts:

[0057] V1-1 Multi-Protocol Adapter Interface: Supports multiple communication protocols, including Ethernet, 4G / 5G, and underwater acoustic communication, ensuring smooth data access for various underwater devices. This interface features automatic protocol identification and conversion, providing a unified access standard for data from different sources.

[0058] V1-2 Data Buffering and Flow Control: Utilizes a distributed message queue to manage data flow and employs intelligent flow regulation to prevent data congestion. The system can smoothly handle data spikes, ensuring stable data transmission and preventing data loss or system overload.

[0059] The V1-3 real-time stream processing engine processes input data in real time, including video frame segmentation and format conversion. The engine dynamically adjusts processing strategies to ensure efficient and real-time data processing, preparing the data for subsequent analysis.

[0060] V1-4 Data Verification and Integrity Assurance: This system ensures the accuracy of data transmission through verification algorithms, providing data retransmission and integrity verification functions. It can detect and repair transmission errors, guaranteeing the integrity and reliability of data from acquisition to storage.

[0061] 10. As a preferred embodiment of the present invention: the distributed storage system in V2 comprises the following parts:

[0062] V2-1 Multi-Tier Storage Architecture: Employs a tiered storage design, storing hot, warm, and cold data on media with different performance levels. This design ensures both efficient data access and optimal control of storage costs.

[0063] V2-2 Data Organization and Management System: Through a rational directory structure and metadata management, massive amounts of data are kept in an orderly manner. The system supports rapid retrieval and location, providing users with an efficient data management experience.

[0064] V2-3 Data Lifecycle Management: Automatically performs data migration, archiving, and cleanup based on data value and usage frequency. The system intelligently manages the entire data lifecycle and optimizes storage resource utilization.

[0065] 11. As a preferred embodiment of the present invention: the data management module in V3 consists of the following parts:

[0066] The V3-1 metadata management system is responsible for managing descriptive information about data and establishing a comprehensive data indexing system. Through unified metadata standards, it enables rapid data retrieval and precise location.

[0067] V3-2 Data Quality Control System: This system performs multi-dimensional assessment and monitoring of data quality to ensure data meets usage standards. It can automatically identify quality issues and provide improvement suggestions.

[0068] V3-3 Data Preprocessing Pipeline: Provides data cleaning, transformation, and standardization functions to improve data quality. The pipeline uses configurable processing flows to meet different data preprocessing needs.

[0069] V3-4 Data Service Bus: The hub for data flow, providing unified data access and exchange services. It enables smooth data flow between modules through standardized interfaces.

[0070] 12. As a preferred embodiment of the present invention: the security and access control module in V4 comprises the following parts:

[0071] V4-1 Multi-Level Access Control System: Establishes a fine-grained permission management mechanism to ensure secure and controllable data access. The system supports role-based and user-based permission allocation to achieve precise access control.

[0072] V4-2 end-to-end data encryption: Provides end-to-end encryption protection for data transmission and storage to prevent data leakage. It employs advanced encryption algorithms and key management schemes to ensure data security.

[0073] V4-3 Security Audit and Monitoring: Monitors system security status in real time and records all security-related events. Through audit logs and risk alerts, it promptly identifies and addresses security issues.

[0074] V4-4 Privacy Protection Mechanism: Employs data anonymization and desensitization technologies to protect user privacy. The system adheres to privacy protection standards to ensure compliant processing of sensitive data.

[0075] 13. As a preferred embodiment of the present invention: the monitoring and maintenance subsystem in V5 consists of the following parts:

[0076] V5-1 Resource Monitoring System: Real-time monitoring of system resource usage, including computing, storage, and network resources. A visual monitoring interface intuitively displays the system's operational status.

[0077] V5-2 Performance Optimization Tool: Provides system performance analysis and optimization functions to help improve system operating efficiency. The tool can identify performance bottlenecks and provide optimization suggestions.

[0078] V5-3 Fault Diagnosis and Recovery: Equipped with fault detection and automatic recovery capabilities, ensuring stable system operation. The system can quickly locate the cause of the fault and perform repair operations.

[0079] V5-4 Operations and Maintenance Management Platform: Integrates operations and maintenance management functions, providing a unified operations and maintenance interface. The platform supports automated operations and maintenance tasks, improving operations and maintenance efficiency.

[0080] 14. As a preferred embodiment of the present invention: the integration and API layer in V6 consists of the following parts:

[0081] V6-1 Unified API Gateway: The unified entry point for the system, providing API management and routing functions. The gateway supports load balancing and access control to ensure the stability of API services.

[0082] V6-2 Data Service Interface: Provides a standardized data access interface that supports various data operations. The interface is designed to be simple and easy to use, facilitating integration with third-party systems.

[0083] V6-3 System Integration Interface: Provides a standard solution for inter-system integration, supporting data exchange and function calls with other systems. The interface adopts open standards to ensure ease of integration.

[0084] V6-4 Developer Support System: Provides developers with comprehensive support resources, including documentation, examples, and tools. Lowers the development barrier through the developer portal and community support.

[0085] 15. As a preferred embodiment of the present invention: the end-to-end target detection model PDFD-DETR based on the Transformer architecture in S3 consists of the following modules:

[0086] M1: Integrating the C2f module into the backbone network reduces the number of parameters required by the backbone network.

[0087] At the same time, ensure the model's ability to extract lesion features;

[0088] M2: A lightweight deformable spatial attention network (DSAN) is introduced into the feature extraction network. Its deformable convolution mechanism enhances the adaptability to irregular lesion areas on the fish body surface and further reduces the number of model parameters.

[0089] M3: The PolaFormer polarity-aware attention mechanism is introduced into the encoding layer to replace the original multi-head self-attention mechanism. This mechanism effectively captures subtle feature differences by decomposing the positive and negative components of features and independently calculating their interactions, further improving the model's ability to distinguish between lesion areas and healthy areas on the fish's body surface;

[0090] M4: Model optimization and training strategies;

[0091] 16. As a preferred embodiment of the present invention: the overall model architecture design in M1 consists of the following parts:

[0092] M1-1 backbone network selection module: Integrates a lightweight CSPDarknet network to replace the traditional ResNet. It optimizes gradient flow through cross-stage local (CSP) design, reduces the number of parameters, and maintains multi-scale feature extraction capability.

[0093] The M1-2 encoding layer construction module includes an attention-based intra-scale feature interaction module (AIFI) and a CNN-based cross-scale feature fusion module (CCFF), which are used to encode and fuse extracted features at multiple scales to improve feature representation capabilities.

[0094] M1-3 Decoding Layer and Query Mechanism Module: Employs an IoU-aware query selection mechanism to provide the decoder with high-quality initial object queries, avoiding NMS post-processing and directly outputting prediction results.

[0095] The M1-4 end-to-end pipeline integration module coordinates the workflow of the backbone network, encoder layer, and decoder layer, ensuring seamless connection of the model from input to output and supporting real-time detection requirements.

[0096] 17. As a preferred embodiment of the present invention: the model optimization and training strategy in M4 consists of the following parts:

[0097] The M4-1 loss function design module combines classification loss (FocalLoss) and regression loss (GIoULoss) to address class imbalance and bounding box accuracy issues, thereby improving detection accuracy.

[0098] M4-2 optimizer configuration module: Uses AdamW optimizer, sets momentum to 0.9, weight decay to 0.0001, and uses cosine annealing scheduling for the learning rate to accelerate model convergence.

[0099] The M4-3 hyperparameter setting module controls batch size and training epochs, and integrates data augmentation strategies to enhance model generalization ability.

[0100] M4-4 Regularization and Overfitting Prevention Module: Employs DropPath and LabelSmoothing techniques to prevent overfitting during training.

[0101] 18. As a preferred embodiment of the present invention: the lesion visualization analysis tool in S4 consists of the following modules:

[0102] N1: Input interface and data receiving module

[0103] N2: Model Inference and Output Engine

[0104] N3: Visualization Rendering and Result Presentation

[0105] N4: Interpretability Analysis Tool

[0106] N5: Real-time monitoring and early warning function

[0107] N6: Decision Support and Report Generation

[0108] 19. As a preferred embodiment of the present invention: the input interface and data receiving module in N1 consist of the following parts:

[0109] N1-1 Multi-Source Data Access Interface: Supports receiving image data from underwater camera equipment, real-time video streams, or storage systems, is compatible with multiple formats, and provides an API interface for data uploading and querying, ensuring the diversity and accessibility of data sources.

[0110] N1-2 Data Parsing and Preprocessing Unit: Performs preliminary parsing of input data, including metadata extraction, format conversion, and size standardization, in preparation for subsequent model inference.

[0111] N1-3 Buffering and Queue Management: Message queues are used to buffer peak data traffic, enabling smooth data inflow, avoiding system overload, and supporting priority scheduling to ensure that critical data is processed first.

[0112] 20. As a preferred embodiment of the present invention: the model inference and result output engine in N2 consists of the following parts:

[0113] N2-1 Model Loading and Configuration Management: Integrates the trained PDFD-DETR model, supports GPU-accelerated inference, automatically loads model weights and configuration files, and handles model version management and updates.

[0114] N2-2 Real-time Inference Engine: Based on the PyTorch or TensorRT framework, it performs end-to-end inference on the input image and outputs the localization and classification information of the disease area. The inference speed meets the real-time requirements.

[0115] N2-3 Result Parsing and Formatting: Convert the model output into structured data, including bounding boxes, class labels, confidence scores, and add timestamps and image identifiers to facilitate subsequent visualization and analysis.

[0116] 21. As a preferred embodiment of the present invention: the visualization rendering and result presentation in N3 consists of the following parts:

[0117] The N3-1 graphics overlay rendering engine overlays prediction results onto the original image, including drawing bounding boxes, labels, and confidence scores. It supports color encoding and provides interactive operations such as scaling, translation, and rotation.

[0118] The N3-2 multi-view display interface provides side-by-side comparison views, displaying the original image, prediction results, and heatmap. It supports split-screen and tab switching, helping users to intuitively compare and analyze the detection results.

[0119] N3-3 Dynamic Interface and Interactive Controls: Integrates a web-based GUI or mobile interface, providing controls such as sliders and drop-down menus for adjusting confidence thresholds, filtering disease types, or viewing historical records, enhancing user engagement.

[0120] 22. As a preferred embodiment of the present invention: the interpretability analysis tool in N4 consists of the following parts:

[0121] N4-1 Heatmap Generator: Integrates Grad-CAM technology to generate category-activated heatmaps, intuitively displaying the model's focus on lesion areas. The heatmaps use color gradients and support transparency adjustment and export functions.

[0122] N4-2 Feature Response Analysis Tool: Visualizes the feature maps of the intermediate layers of the model, showcasing the feature extraction effects of the DSAN Block and PolaFormer modules, and helping users understand how the model captures irregular lesion areas.

[0123] N4-3 Confidence Distribution Visualization: Generates bar charts, pie charts, or line graphs to display the confidence distribution and statistical information of various diseases, supporting interactive queries and filtering.

[0124] 23. As a preferred embodiment of the present invention: the real-time monitoring and early warning function in the N5 consists of the following parts:

[0125] N5-1: Real-time stream processing pipeline: Connects to video stream input, analyzes frame by frame and dynamically updates detection results, with a frame rate of no less than 25fps to ensure real-time performance.

[0126] N5-2: Automatic Alarm Trigger: Set confidence thresholds and rules engine to trigger sound, email, or SMS alerts and push location information and suggested actions.

[0127] N5-3: Historical Data Tracker: Records time-series data of detection results, generates trend charts, supports backtracking queries and anomaly detection, and assists in long-term monitoring.

[0128] 24. As a preferred embodiment of the present invention: the decision support and report generation in N6 comprises the following parts:

[0129] N6-1: Expert Knowledge Base Integration: Built-in fish disease treatment suggestion library, automatically recommends countermeasures based on test results, and supports custom rules.

[0130] N6-2: One-click report generator: Export PDF or HTML reports, including a summary of detection results, visualizations, trend analysis, and processing suggestions. It supports template customization and batch generation.

[0131] N6-3: Data Export and Sharing Interface: Provides data export in formats such as JSON and CSV, facilitating integration into third-party systems, and supports data sharing and collaboration through API. Attached Figure Description

[0132] Figure 1 This is a flowchart illustrating the overall architecture of the method of the present invention.

[0133] Figure 2 This is a diagram of the overall system architecture of the present invention.

[0134] Figure 3 This is a structural diagram of the PDFD-DETR of the present invention.

[0135] Figure 4 This is a structural diagram of the C2f_DSAN of the present invention.

[0136] Figure 5 This is a structural diagram of the AIFI-Pola of the present invention.

[0137] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A lightweight method and system for detecting fish surface diseases based on improved RT-DETR, characterized in that, Specifically, the following steps are included: S1: Use underwater camera equipment to collect image data of fish body surface, covering a variety of common fish body surface diseases, including but not limited to ulcers, gill rot, fin damage, eye diseases, etc. S2: The collected data is stored and managed through a high-performance data processing platform to ensure data integrity and accessibility; S3: Construct an end-to-end object detection model PDFD-DETR based on the Transformer architecture. Integrate the lightweight variable convolution module DSANBlock and the PolaFormer polarity-aware attention mechanism into the RT-DETR model to enhance the model's adaptability to irregular lesion areas on the fish body surface and its ability to capture lesion details. S4: A lesion visualization and analysis tool that inputs images of fish to be detected into the system and outputs location and classification information of disease areas, enhancing the interpretability of the results, supporting real-time monitoring and decision-making applications, and improving interactivity and comprehensibility.

2. The lightweight fish surface disease detection method and system based on improved RT-DETR according to claim 1, characterized in that: The collection of underwater fish health status data in S1 is composed of the following modules of the system: A: Underwater high-definition camera equipment; B: Underwater auxiliary lighting system; C: Waterproof housing and pressure protection; D: Underwater support and stabilization device; E: Control and monitoring system.

3. The lightweight fish surface disease detection method and system based on improved RT-DETR according to claim 1, characterized in that: The high-efficiency data processing in S2 consists of the following modules composition: V1: Data receiving and access layer; V2: Distributed storage system; V3: Data Management Module; V4: Security and Access Control Module; V5: Monitoring and Maintenance Subsystem; V6: Integration and API Layer.

4. The lightweight fish surface disease detection method and system based on improved RT-DETR according to claim 1, characterized in that: The end-to-end object detection model PDFD-DETR based on the Transformer architecture in S3 consists of the following modules: M1: Integrating the C2f module into the backbone network reduces the number of parameters required by the backbone network. At the same time, ensure the model's ability to extract lesion features; M2: A lightweight deformable spatial attention network (DSAN) is introduced into the feature extraction network. Its deformable convolution mechanism enhances the adaptability to irregular lesion areas on the fish body surface and further reduces the number of model parameters. M3: The PolaFormer polarity-aware attention mechanism is introduced into the encoding layer to replace the original multi-head self-attention mechanism. This mechanism effectively captures subtle feature differences by decomposing the positive and negative components of features and independently calculating their interactions, further improving the model's ability to distinguish between lesion areas and healthy areas on the fish's body surface; M4: Model optimization and training strategies.

5. The lightweight fish surface disease detection method and system based on improved RT-DETR according to claim 1, characterized in that: The lesion visualization and analysis tool in S4 consists of the following modules composition: N1: Input interface and data receiving module; N2: Model inference and result output engine; N3: Visualization rendering and result presentation; N4: Interpretability analysis tool; N5: Real-time monitoring and early warning function; N6: Decision support and report generation.

Images