An intelligent detection system of RGB camera combined with deep learning algorithm

Abstract

The application provides an intelligent detection system of an RGB camera combined with a deep learning algorithm, comprising a defect feature adaptive engine, a hardware component, an algorithm component and a feedback component; the defect feature adaptive engine serves as a core control unit, links the hardware component, the algorithm component and the feedback component to realize full-link closed-loop detection; the defect feature adaptive engine performs the following operations: identifying scene working conditions and core working condition pain points, extracting the subdivided features of target defects in the scene and matching the feature label library, dynamically calling the adaptive hardware configuration, algorithm combination and preprocessing strategy, optimizing the detection parameters in real time, allocating the computing power resources according to the defect risk level and triggering the corresponding early warning, realizing the intelligent identification, positioning, quantification and grading early warning of defects of equipment such as pressure pipelines and pressure vessels, and through the implementation of the above system, the problems of poor adaptability, insufficient precision and unreasonable resource allocation of the existing system are solved, and the industrial detection is realized in a scene-based, refined and efficient manner.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for industrial equipment, specifically to an intelligent inspection system using an RGB camera and a deep learning algorithm. Background Technology

[0002] With the long-term operation and diversified working conditions of industrial equipment, equipment defects in different scenarios exhibit significant differences in their detailed characteristics, and the requirements for the scenario adaptability of detection systems are increasing.

[0003] Traditional manual inspection suffers from low efficiency, high safety risks, and a high rate of missed detections of minor defects, making it unable to meet the inspection needs of complex industrial scenarios. While existing intelligent inspection systems using RGB cameras and deep learning achieve automated inspection, they have several key limitations: First, defect features are not segmented according to specific scenarios, and a generalized algorithm is used to adapt to all scenarios, resulting in significant differences in the accuracy of detecting the same defect in different scenarios. Second, the potential for scenario-specific applications of document-specific technologies is not fully explored; for example, the invasive and multi-angle observations by robotic arms are not adapted to different scenarios, and adaptive preprocessing strategies are not dynamically adjusted to address specific operational challenges. Third, the detection priority is unclear, using the same computing power for high-risk and low-risk defects, leading to resource waste. Fourth, there is a lack of specific optimization strategies for small-sample scenarios, resulting in insufficient model generalization ability.

[0004] Therefore, there is an urgent need to design an intelligent detection system based on scenario-specific defect feature adaptation. By accurately binding scenario, defect subdivision features, and technical solutions, this system can solve the problems of poor adaptability, insufficient accuracy, and unreasonable resource allocation in existing systems, and realize the scenario-based, refined, and efficient implementation of industrial inspection. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems mentioned above and to propose an intelligent detection system using an RGB camera and a deep learning algorithm, including a defect feature adaptive engine, hardware components, algorithm components, and feedback components. The defect feature adaptive engine serves as the core control unit, linking hardware components, algorithm components, and feedback components to achieve full-link closed-loop detection. The defect feature adaptive engine performs the following operations: identifies the scene conditions and core pain points, extracts the detailed features of the target defect in the scene and matches them with the feature tag library, dynamically calls the appropriate hardware configuration, algorithm combination and preprocessing strategy, optimizes the detection parameters in real time, allocates computing resources according to the defect risk level and triggers corresponding warnings, so as to realize the intelligent identification, location, quantification and graded warning of defects in equipment such as pressure pipelines and pressure vessels.

[0006] In the preferred embodiment, the defect feature adaptive engine includes a scene condition recognition module. The scene condition recognition module automatically identifies at least one scene type among large hydropower stations, chemical acid and alkali, marine / coastal, wind power, high-temperature boilers, and rail transit by using image features collected by RGB cameras and data collected by auxiliary sensors, and labels the corresponding pain points of the scene condition.

[0007] In the preferred embodiment, the defect feature adaptive engine includes a defect subdivision feature extraction module. Based on the scene recognition results, the defect subdivision feature extraction module automatically extracts subdivision features of at least one defect among bolt loosening, corrosion, cracks, deformation, and paint peeling. The subdivision features include pitting corrosion, creep intergranular cracks, rust blistering peeling, and small gap loosening, forming a scene-defect feature tag library.

[0008] In the preferred embodiment, the defect feature adaptive engine includes a technical component matching module. Based on the feature tag library, the technical component matching module automatically links and calls the working mode of the hardware components, the combination method of the algorithm components, and the preprocessing strategy to achieve accurate adaptation of the scenario, defect, and technical solution.

[0009] In the preferred embodiment, the defect feature adaptive engine includes a dynamic optimization module. The dynamic optimization module provides real-time feedback on the detection results. If a missed detection or false detection occurs, it automatically adjusts the algorithm parameters, preprocessing intensity, and hardware working mode until the preset detection accuracy requirements are met.

[0010] In the preferred embodiment, the hardware components include a high-resolution RGB camera, a multi-degree-of-freedom robotic arm, and auxiliary sensors; the RGB camera can switch between low-light mode, anti-corrosion mode, and high-temperature mode under engine control; the robotic arm can switch between invasive or multi-angle inspection mode under engine control; the auxiliary sensors include at least one of lidar, temperature and humidity sensors, and vibration sensors, which are selectively used by the engine according to the detection requirements.

[0011] In the preferred embodiment, the algorithm components include a core algorithm library, an optimization algorithm library, and a preprocessing algorithm library; the core algorithm library includes the YOLOv8 object detection algorithm, the U-Net / U-Net++ semantic segmentation algorithm, and the Sobel edge detection algorithm, which are combined and called by the engine according to the defect localization or quantization requirements; The optimization algorithm library includes attention mechanisms, feature pyramid fusion, GAN generative models, and transfer learning algorithms, which are selectively called by the engine based on defect features; The preprocessing algorithm library includes CLAHE illumination compensation, polarization filtering, salt spray filtering, and vibration denoising algorithms, which are matched and called by the engine according to the pain points of the working conditions.

[0012] In the preferred scheme, the defect feature adaptive engine has a built-in small sample optimization module. After the small sample optimization module identifies small sample scenarios such as high-temperature creep cracks and marine pitting, it calls the GAN generative model to generate scenario-specific virtual samples. The transfer learning algorithm is then used to integrate general model features with virtual samples and a small number of real samples for training, thereby optimizing the model's generalization ability.

[0013] In the preferred embodiment, the defect feature adaptive engine includes a risk classification and scheduling module. The risk classification and scheduling module classifies cracks and severe deformation into high-risk defects and allocates computing resources for dual-algorithm verification and multi-sensor fusion; it classifies loose bolts and large-area corrosion into medium-risk defects and allocates computing resources for conventional algorithms and single-sensor verification; and it classifies minor paint peeling and uniform surface rust into low-risk defects and allocates lightweight algorithm configurations to achieve dynamic allocation of computing power.

[0014] In the preferred embodiment, the feedback component includes a hierarchical alarm mechanism. Under engine control, the hierarchical alarm mechanism triggers a first-level real-time push alarm for high-risk defects, a second-level data storage and trend tracking alarm for medium-risk defects, and a third-level routine record alarm for low-risk defects, while simultaneously outputting the defect type, location, quantitative parameters, and risk assessment report.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Through the precise matching of “scene conditions - defect subdivision features - technical components” by the defect feature adaptive engine, the problem of differences in subdivision features of the same defect under different scenarios is effectively solved, and the missed detection and false detection caused by working conditions such as salt spray reflection, vibration blurring, and low light are greatly reduced. The detection effect stability is significantly better than that of general-purpose detection systems. 2) The engine automatically completes scene recognition, linkage invocation of technical components, and dynamic parameter optimization, without the need for manual adjustment of hardware mode, algorithm parameters, or preprocessing strategies, enabling rapid switching and adaptation across multiple scenes and significantly reducing debugging costs and time for cross-scene deployment. 3) Through the engine's risk-based scheduling mechanism, high-risk defects receive computing power allocation through dual-algorithm verification and multi-sensor fusion, while low-risk defects are configured with lightweight algorithms to avoid wasting computing power and to balance detection real-time performance and accuracy. 4) By linking GAN generative samples and transfer learning, the feature capture capability of small sample defects such as high temperature creep cracks and marine pitting is specifically enhanced. Ideal detection results can be achieved without training with a large number of real samples, which greatly reduces the cost of sample collection and annotation. 5) The engine is based on existing mature technology components and achieves performance breakthroughs through innovative linkage design. There is no need to develop additional dedicated hardware or algorithms. At the same time, it reduces manual intervention, lowers system deployment, debugging and maintenance costs, and is easier to promote and apply in industrial sites. 6) The defect information output by the engine includes type, location, quantitative parameters and risk level assessment. Combined with operating data, it forms a complete inspection report, providing accurate basis for preventive maintenance of equipment and helping to reduce equipment failure risk and maintenance costs. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the contrast enhancement effect.

[0017] Figure 2 This is a flowchart for image edge enhancement.

[0018] Figure 3 It is the YOLOv8 neural network architecture.

[0019] Figure 4 This is a schematic diagram of YOLOv8 performing bolt inspection.

[0020] Figure 5 It is the UNet neural network architecture.

[0021] Figure 6 This is a schematic diagram of semantic segmentation of the rusted area.

[0022] Figure 7 This is a schematic diagram of semantic segmentation for crack detection.

[0023] Figure 8 This is a flowchart of a multi-level defect detection process. Detailed Implementation

[0024] Example 1 This embodiment provides an intelligent detection system using an RGB camera and a deep learning algorithm, including a defect feature adaptive engine, hardware components, algorithm components, and feedback components; The defect feature adaptive engine serves as the core control unit, linking hardware components, algorithm components, and feedback components to achieve full-link closed-loop detection. The defect feature adaptive engine performs the following operations: identifies the scene conditions and core pain points, extracts the detailed features of the target defect in the scene and matches them with the feature tag library, dynamically calls the appropriate hardware configuration, algorithm combination and preprocessing strategy, optimizes the detection parameters in real time, allocates computing resources according to the defect risk level and triggers corresponding warnings, so as to realize the intelligent identification, location, quantification and graded warning of defects in equipment such as pressure pipelines and pressure vessels.

[0025] Preferably, the defect feature adaptive engine includes a scene condition recognition module. The scene condition recognition module automatically identifies at least one scene type among large hydropower stations, chemical acid and alkali, marine / coastal, wind power, high-temperature boilers, and rail transit by using image features collected by an RGB camera and data collected by auxiliary sensors, and labels the corresponding pain point tags.

[0026] Preferably, the defect feature adaptive engine includes a defect subdivision feature extraction module. Based on the scene recognition results, the defect subdivision feature extraction module automatically extracts subdivision features of at least one defect among bolt loosening, rust, cracks, deformation, and paint peeling. The subdivision features include pitting, creep intergranular cracks, rust blistering, and small gap loosening, forming a scene-defect feature tag library.

[0027] Preferably, the defect feature adaptive engine includes a technical component matching module. The technical component matching module automatically links and calls the working mode of hardware components, the combination method of algorithm components, and the preprocessing strategy according to the feature tag library to achieve accurate matching of scenario-defect-technical solution.

[0028] Preferably, the defect feature adaptive engine includes a dynamic optimization module, which provides real-time feedback on the detection results. If a missed detection or false detection occurs, the algorithm parameters, preprocessing intensity, and hardware working mode are automatically adjusted until the preset detection accuracy requirements are met.

[0029] Preferably, the hardware components include a high-resolution RGB camera, a multi-degree-of-freedom robotic arm, and auxiliary sensors; the RGB camera can switch between low-light mode, anti-corrosion mode, and high-temperature mode under engine control; the robotic arm can switch between invasive or multi-angle inspection mode under engine control; the auxiliary sensors include at least one of lidar, temperature and humidity sensor, and vibration sensor, which are selectively called by the engine according to the detection requirements.

[0030] Preferably, the algorithm components include a core algorithm library, an optimization algorithm library, and a preprocessing algorithm library; the core algorithm library includes the YOLOv8 object detection algorithm, the U-Net / U-Net++ semantic segmentation algorithm, and the Sobel edge detection algorithm, which are combined and called by the engine according to the defect localization or quantization requirements; The optimization algorithm library includes attention mechanisms, feature pyramid fusion, GAN generative models, and transfer learning algorithms, which are selectively called by the engine based on defect features; The preprocessing algorithm library includes CLAHE illumination compensation, polarization filtering, salt spray filtering, and vibration denoising algorithms, which are matched and called by the engine according to the pain points of the working conditions.

[0031] Preferably, the defect feature adaptive engine has a built-in small sample optimization module. After identifying small sample scenarios such as high-temperature creep cracks and marine pitting, the small sample optimization module calls the GAN generative model to generate scenario-specific virtual samples. The transfer learning algorithm is then used to integrate general model features with virtual samples and a small number of real samples for training, thereby optimizing the model's generalization ability.

[0032] Preferably, the defect feature adaptive engine includes a risk classification and scheduling module. The risk classification and scheduling module classifies cracks and severe deformation into high-risk defects and allocates computing resources for dual-algorithm verification and multi-sensor fusion; it classifies loose bolts and large-area corrosion into medium-risk defects and allocates computing resources for conventional algorithms and single-sensor verification; and it classifies slight paint peeling and uniform surface rust into low-risk defects and allocates lightweight algorithm configurations to achieve dynamic allocation of computing power.

[0033] Preferably, the feedback component includes a hierarchical alarm mechanism. Under engine control, the hierarchical alarm mechanism triggers a level 1 real-time push alarm for high-risk defects, a level 2 data storage and trend tracking alarm for medium-risk defects, and a level 3 routine record alarm for low-risk defects, while simultaneously outputting the defect type, location, quantitative parameters, and risk assessment report.

[0034] The solution in this embodiment is specifically applicable to the intelligent identification, location, quantification, and graded early warning of defects such as loose bolts, corrosion, cracks, deformation, and paint peeling in pressure pipelines and pressure vessels in various scenarios including large-scale hydropower stations, chemical acid and alkali plants, marine / coastal plants, wind power, high-temperature boilers, and rail transportation. The core innovation lies in achieving dynamic and accurate matching of scenario, defect, and technical solution through a "defect feature adaptive engine," thus solving the problems of poor generalization and insufficient accuracy in existing systems.

[0035] The mapping rules for scenario-defect-technical component are shown in Table 1: Table 1. Scenario-Defect-Technical Component Mapping Rules

[0036] Taking pitting detection in marine / coastal scenarios as an example, the engine-linked full flow is as follows: 1. Scene and working condition recognition: The RGB camera captures images of the pipe surface, and the engine detects the presence of "salt spray coverage features + strong reflective signals" in the image. Combined with temperature and humidity sensor data, it automatically identifies the scene as "ocean / coastal scene" and the working condition pain point label is "salt spray + reflective". 2. Defect Feature Extraction: The engine performs a preliminary analysis of the rusted areas in the image and extracts the features of "dense small pits + crevice corrosion". By matching the feature label library, the defect type is determined to be "marine scene pitting corrosion". 3. Technical component invocation: The engine automatically invokes the following combinations: Hardware: The corrosion-resistant RGB camera is switched to "salt spray mode", the polarizing filter is activated to filter reflections, and the temperature and humidity sensor is used to collect data synchronously; Preprocessing: The salt spray filtering algorithm is activated to remove the haze from the image, polarization filtering is used to eliminate reflection interference, and super-resolution processing is performed on the pitted area to magnify the details; Algorithm: Call the "pitting quantization branch" of U-Net semantic segmentation, and superimpose the attention mechanism to focus on the pitting region, accurately segment the pits and calculate the depth / area; 4. Dynamic optimization: The engine monitors the detection results in real time. If some small pits are missed, it automatically increases the CLAHE enhancement intensity and adjusts the U-Net segmentation threshold until the detection accuracy requirements are met. 5. Risk classification output: The engine automatically triggers a level 2 alarm based on the pit depth (≥3mm), synchronously stores the detection data and associates it with temperature and humidity sensors to analyze the corrosion rate trend.

[0037] The implementation process for other scenarios follows the logic of "engine recognition → feature extraction → component invocation → dynamic optimization → hierarchical output". Only the combination of technical components is automatically adjusted by the engine according to the differences in scenario-defect characteristics, without the need for manual intervention.

[0038] Example 2 like Figures 1-8 As shown, this embodiment provides an intelligent inspection system using an RGB camera and a deep learning algorithm, relating to the field of industrial inspection. Addressing the difficulties and low efficiency of traditional inspection methods in complex environments, this invention combines a high-resolution RGB camera with a deep learning algorithm to capture and process image information from equipment surfaces in real time. The system utilizes the YOLOv8 target detection model and semantic segmentation technology to achieve intelligent identification and location of various defects such as pipes, loose bolts, deformed supports, and cracks. Image preprocessing techniques, such as noise reduction and edge enhancement, ensure high-precision inspection in complex industrial environments. Furthermore, the system's multi-degree-of-freedom robotic arm design allows for flexible adjustment of the camera angle, enabling in-depth and multi-angle observations to ensure full inspection coverage. This system not only improves inspection accuracy but also enhances maintenance efficiency through a real-time alarm mechanism, providing strong protection for the safe operation of industrial equipment. Specific implementation details are as follows: 1. Application of RGB cameras combined with deep learning algorithms 1.1 Image Acquisition and Preprocessing This intelligent inspection system uses a high-performance RGB camera to acquire real-time image data of the target inspection area. The camera acquires high-resolution real-time images of the surface of equipment such as factory pipelines and pressure vessels, and then performs subsequent processing using deep learning algorithms. To improve image quality, factors such as lighting and color are automatically adjusted during image acquisition to ensure clear image information is obtained even in complex environments, such as low light or high-contrast backgrounds.

[0039] 1.2 Applications of Deep Learning Algorithms After preprocessing, image data is input into a deep learning module for analysis. Advanced deep learning models, such as Convolutional Neural Networks (CNNs), are trained and used to identify surface defects in equipment within the images. Trained on a large amount of industrial sample data, the deep learning model can automatically identify common defects such as cracks, rust, corrosion, and coating peeling, and accurately locate the defects based on their position and features in the image. This deep learning algorithm can analyze and generate defect reports in real time, assisting maintenance personnel in quickly identifying potential equipment malfunctions.

[0040] 1.3 Automatic Defect Detection and Feedback Mechanism By combining image data acquired by an RGB camera with the automatic recognition capabilities of deep learning algorithms, the system can achieve a fully automated detection mode. During the detection process, the system automatically generates information such as the type, location, and size of defects based on image analysis results. All detection results are fed back to the operator in real time through the control system, ensuring timely handling of potential problems. This automated process significantly improves detection efficiency and accuracy.

[0041] 2. Implementation of the Deep Learning Defect Identification and Analysis Module 2.1 Target Detection and Classification The deep learning module analyzes RGB images using object detection algorithms and classifies detected targets (such as cracks, rust, and coating peeling on equipment surfaces). During implementation, the object detection module employs feature extraction and image classification techniques to accurately distinguish different types of defects and calculate the relative position and size of each defect. This classification function enables the system to provide diagnostic reports for different types of defects, facilitating subsequent equipment maintenance and repair.

[0042] 2.2 Semantic Segmentation and Precise Localization To ensure detection accuracy, this system employs semantic segmentation technology, dividing the device image into multiple regions to more accurately identify and locate defects in each region. Deep learning algorithms perform pixel-by-pixel analysis of each region, automatically labeling defect locations and outputting the defect's shape, size, and relative position to other regions. This technology effectively handles small and minute defects against complex backgrounds, ensuring accurate detection even in extremely complex working environments.

[0043] 2.3 Data Storage and Subsequent Analysis All inspection data, including defect type, inspection location, and dimensions, is stored in the database. The system automatically generates a detailed inspection report for each inspection and compares it with historical data to help maintenance personnel understand the changing trends of equipment status, thereby enabling better preventative maintenance. This data can also be analyzed by technicians through a remote monitoring platform, providing support for future equipment management.

[0044] 3. System Architecture and Automated Control 3.1 Design and Functions of the Intelligent Control Module The core control modules of this system include an image acquisition module, a data processing module, and an execution control module. The image acquisition module is responsible for acquiring real-time images of the equipment surface. The data processing module performs preprocessing and deep learning analysis on the images. The execution control module executes necessary operations based on the analysis results, such as automatic report generation and data transmission. The control module can adjust the working mode (automatic or manual control) of each part according to on-site requirements.

[0045] 3.2 Automated Detection Path Planning During the inspection process, this system automatically determines the optimal inspection path through a path planning algorithm, ensuring coverage of the entire inspection area. The system can flexibly adjust the inspection path according to the complexity of the on-site environment and automatically avoid obstacles using autonomous navigation technology, ensuring the smooth progress of the inspection process. The robot's motion path planning in complex environments is supported by the autonomous navigation system, enabling it to complete comprehensive inspections without interfering with the equipment.

[0046] 3.3 Remote Monitoring and Operation This system connects to a remote monitoring platform via a wireless communication module. Operators can view the equipment's testing status, historical testing data, and reports in real time through the monitoring platform, and schedule or manually intervene in testing tasks as needed. In manual operation mode, operators can make precise adjustments through the control interface to ensure the system can successfully complete tasks even in special circumstances.

[0047] 4. System Hardware and Software Implementation 4.1 Selection and Integration of Hardware Components The hardware configuration of this intelligent detection system mainly includes a high-resolution RGB camera, a deep learning computing unit, a LiDAR (such as a distance sensor suitable for complex environments), and necessary sensors (such as temperature and humidity sensors). The RGB camera was selected based on high resolution, adaptability to low-light environments, and high frame rate requirements to ensure clear and stable images can be obtained under various environmental conditions.

[0048] 4.2 Software Platform and Algorithm Optimization The software platform uses deep learning frameworks (such as TensorFlow or PyTorch) for algorithm implementation and optimization. All deep learning models and data processing algorithms are trained and inferred via cloud or edge computing devices to ensure real-time performance and accuracy. The algorithm optimization module regularly updates and optimizes the detection model to improve detection accuracy and operational efficiency, adapting to the detection needs of different types of devices.

[0049] 4.3 Multi-level data fusion The system can perform comprehensive analysis by combining RGB camera image data, LiDAR data, and data from other sensors using multi-sensor data fusion technology. This multi-sensor fusion scheme enhances the reliability of detection, enabling the system to make accurate judgments about defects in complex environments.

[0050] 5. Safety and stability design 5.1 Redundancy Design and Fault Tolerance To enhance system stability and reliability, this system employs a redundant design, ensuring automatic switching to a backup scheme to maintain continuous operation in the event of a failure in individual sensors or modules. Control modules and critical components are all designed with fault tolerance to address potential risks under high-pressure operating environments.

[0051] 5.2 System Protection and Security Assurance To ensure long-term stable operation in industrial environments, the system features a high level of hardware protection, including dust and water resistance, and the ability to withstand common industrial vibrations and electromagnetic interference. The software incorporates multi-level security measures, such as user authentication and operation log recording, to ensure compliance with safety regulations during operation and prevent misoperation or malicious attacks.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent detection system using an RGB camera and a deep learning algorithm, characterized in that, This includes a defect feature adaptive engine, hardware components, algorithm components, and feedback components; The defect feature adaptive engine serves as the core control unit, linking hardware components, algorithm components, and feedback components to achieve full-link closed-loop detection. The defect feature adaptive engine performs the following operations: identifies the scene conditions and core pain points, extracts the detailed features of the target defect in the scene and matches them with the feature tag library, dynamically calls the appropriate hardware configuration, algorithm combination and preprocessing strategy, optimizes the detection parameters in real time, allocates computing resources according to the defect risk level and triggers corresponding warnings, so as to realize the intelligent identification, location, quantification and graded warning of defects in equipment such as pressure pipelines and pressure vessels.

2. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The defect feature adaptive engine includes a scene condition recognition module. This module uses image features collected by an RGB camera and data collected by auxiliary sensors to automatically identify at least one scene type among large hydropower stations, chemical acid and alkali, marine / coastal, wind power, high-temperature boilers, and rail transit, and labels the corresponding pain point tags.

3. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The defect feature adaptive engine includes a defect subdivision feature extraction module. Based on the scene recognition results, the defect subdivision feature extraction module automatically extracts subdivision features of at least one defect among bolt loosening, rust, cracks, deformation, and paint peeling. Subdivision features include pitting, creep intergranular cracks, rust blistering, and small gap loosening, forming a scene-defect feature label library.

4. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The defect feature adaptive engine includes a technical component matching module. Based on the feature tag library, the technical component matching module automatically links and calls the working mode of hardware components, the combination method of algorithm components, and the preprocessing strategy to achieve accurate adaptation of scenario-defect-technical solution.

5. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The defect feature adaptive engine includes a dynamic optimization module, which provides real-time feedback on the detection results. If a missed detection or false detection occurs, it automatically adjusts the algorithm parameters, preprocessing intensity, and hardware working mode until the preset detection accuracy requirements are met.

6. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The hardware components include a high-resolution RGB camera, a multi-degree-of-freedom robotic arm, and auxiliary sensors. The RGB camera can switch between low-light mode, anti-corrosion mode, and high-temperature mode under engine control. The robotic arm can switch between invasive or multi-angle inspection mode under engine control. The auxiliary sensors include at least one of lidar, temperature and humidity sensors, and vibration sensors, which are selectively used by the engine according to the detection requirements.

7. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The algorithm components include a core algorithm library, an optimization algorithm library, and a preprocessing algorithm library. The core algorithm library includes the YOLOv8 object detection algorithm, the U-Net / U-Net++ semantic segmentation algorithm, and the Sobel edge detection algorithm, which are combined and called by the engine according to the defect localization or quantization requirements. The optimization algorithm library includes attention mechanisms, feature pyramid fusion, GAN generative models, and transfer learning algorithms, which are selectively called by the engine based on defect features; The preprocessing algorithm library includes CLAHE illumination compensation, polarization filtering, salt spray filtering, and vibration denoising algorithms, which are matched and called by the engine according to the pain points of the working conditions.

8. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The defect feature adaptive engine has a built-in few-sample optimization module. After identifying few-sample scenarios such as high-temperature creep cracks and marine pitting, the few-sample optimization module calls the GAN generative model to generate scenario-specific virtual samples. The transfer learning algorithm is then used to integrate general model features with virtual samples and a small number of real samples for training, thereby optimizing the model's generalization ability.

9. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The defect feature adaptive engine includes a risk classification and scheduling module. This module classifies cracks and severe deformation as high-risk defects and allocates computing resources for dual-algorithm verification and multi-sensor fusion. It classifies loose bolts and large-area corrosion as medium-risk defects and allocates computing resources for conventional algorithms and single-sensor verification. It classifies minor paint peeling and uniform surface rust as low-risk defects and allocates lightweight algorithm configurations to achieve dynamic allocation of computing power.

10. The intelligent detection system using an RGB camera and a deep learning algorithm according to claim 1, characterized in that, The feedback component includes a tiered alarm mechanism. Under engine control, the tiered alarm mechanism triggers a level 1 real-time push alarm for high-risk defects, a level 2 data storage and trend tracking alarm for medium-risk defects, and a level 3 routine record alarm for low-risk defects. It also outputs the defect type, location, quantitative parameters, and risk assessment report simultaneously.

Images