Real-time image enhancement and intelligent exposure method and system for unmanned aerial vehicle inspection

Abstract

The invention relates to an unmanned aerial vehicle inspection real-time image enhancement and exposure method and system, and the method comprises the steps: obtaining a real-time original video stream captured by an airborne camera of an unmanned aerial vehicle, enabling an airborne edge computing platform to continuously carry out the light-weight dark light enhancement algorithm on the video stream in real time, and carrying out the basic image preprocessing of the video stream; operating a lightweight target detection model on the preprocessed and enhanced video stream, identifying and positioning a preset inspection target, and accurately marking and outputting the position coordinates and coverage area range of the region of interest of the target; and feeding back the identified position coordinates of the region of interest to a control system of the camera of the unmanned aerial vehicle in real time, enabling the photometric weight of an automatic exposure algorithm of the camera control system to be higher than the region of interest, dynamically adjusting exposure parameters, and capturing an image with clear details. Compared with the prior art, it is ensured that the key target obtains the optimal exposure at the imaging source, and the robustness and recognition accuracy of the downstream model are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of UAV visual processing technology, and in particular to a method and system for real-time image enhancement and intelligent exposure of UAVs for nighttime inspections based on edge AI. Background Technology

[0002] In the field of drone nighttime inspection, high-quality raw image acquisition is a prerequisite for ensuring the accuracy of fault identification and defect detection. Current mainstream nighttime inspection solutions typically rely on large-aperture cameras or high-power supplementary lights. However, under the limited payload of drones and complex nighttime lighting conditions, traditional imaging methods often struggle to balance exposure accuracy and image clarity. Because the nighttime environment is extremely dark, and inspection targets (such as insulators and fasteners) are usually small with varying reflective properties, the camera's automatic exposure (AE) system is prone to metering errors, resulting in motion blur, high noise, or overexposure / underexposure of targets, severely impacting subsequent AI recognition performance. Furthermore, existing edge devices have limited computing power, making it difficult to support complex image restoration algorithms. Therefore, there is an urgent need for an efficient method that combines real-time edge AI perception with camera hardware control.

[0003] For example, Chinese patent application CN202510472324.3 discloses a multimodal fusion method and system for target detection in UAV remote sensing images. This method enhances low-light details by fusing features from infrared and visible light images and utilizing the Retinex principle, aiming to improve detection accuracy under low-light conditions. Chinese patent application CN202510178406.7 discloses a low-light image enhancement method based on a multi-scale fusion U-Net network structure. This method constructs a low-light enhancement network model, preprocesses image data, performs network inference, outputs an enhanced normal-light image, and quantizes the model for deployment in embedded systems. Chinese patent publication CN202410564698.3 discloses a helmet detection method for low-light, dusty environments in coal mines. It uses a parameter generation module to guide adaptive Gamma and white balance algorithms for image enhancement, which is then input into a lightweight detection model. Chinese patent application CN202511004428.8 discloses a lightweight low-light image enhancement system and method. Low-light image sharpening was achieved on the embedded device through a feature extraction module and an illumination channel attention module.

[0004] Existing drone night vision technologies primarily focus on image post-processing, with their core logic lying in using algorithms to repair low-quality images already captured (such as noise reduction, brightening, and multimodal complementation). However, for computer vision, the quality of the original image directly determines the information ceiling. If improper exposure strategies during shooting lead to physical motion blur or pixel overflow (overexposure / underexposure), post-processing enhancement algorithms often struggle to restore true physical details and are prone to introducing artifacts. Existing automatic exposure (AE) technologies are mostly based on global or center metering, failing to detect key regions of interest (ROIs) in the image, resulting in the inability to accurately expose small inspection targets against a large, dark background at night.

[0005] Therefore, how to use AI algorithms to actively guide camera hardware imaging on the drone side with limited computing power, so as to achieve the leap from passive repair to active perception imaging and solve the problems of image clarity and exposure accuracy in nighttime dynamic inspection, has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of motion blur, high noise and low signal-to-noise ratio caused by low light environment in existing drone night inspection technology, and to provide a method and system for real-time image enhancement and exposure of drone night inspection.

[0007] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, a method for real-time image enhancement and exposure during unmanned aerial vehicle (UAV) inspection is provided, comprising the following steps: The system acquires real-time raw video streams captured by the drone's onboard camera. The onboard edge computing platform runs a lightweight low-light enhancement algorithm to enhance the image of the raw video stream, improving the visibility and signal-to-noise ratio of the image. For the enhanced video stream, the airborne edge computing platform runs a lightweight target detection model to identify and locate the preset inspection targets, and marks and outputs the location coordinates and coverage area of ​​the target's region of interest. The location coordinates of the identified region of interest are fed back to the control system of the drone camera in real time. The camera control system then focuses the metering weight of its automatic exposure algorithm on the region of interest and dynamically adjusts the exposure parameters to capture images with clear details.

[0008] As a preferred technical solution, the dark light enhancement algorithm adopts a zero-reference learning model, which uses a set of task-related, reference-free loss functions to guide network learning during the training process.

[0009] As a preferred technical solution, the target detection model adopts a single-stage detector or a hybrid detector based on Transformer.

[0010] As a preferred technical solution, the metering weights of the automatic exposure algorithm are concentrated on the region of interest, specifically as follows: A dynamic metering weighted mask is constructed. Based on the position coordinates of the target's region of interest, a weight matrix with the same resolution as the current video frame is generated in real time. The pixel weights in the region of interest are set to high values, while the background pixel weights outside the region of interest are set to extremely low values. The weighted average of the original image data is calculated using the weight matrix to obtain the weighted brightness value.

[0011] As a preferred technical solution, the dynamic adjustment of exposure parameters is as follows: The maximum allowable exposure time of the shutter is calculated in real time, and the exposure difference is calculated based on the weighted brightness value of the target area and the preset target brightness value. Brightness can be increased by extending the shutter speed, provided that the maximum allowable exposure time is not exceeded. If the image is still underexposed after the shutter speed reaches the maximum allowable exposure time, the required ISO gain is calculated and numerical compensation is performed to increase the image brightness. The calculated shutter speed and ISO gain values ​​are written into the hardware register through the camera control interface to control the acquisition of the next frame image.

[0012] As a preferred technical solution, the dark light enhancement algorithm and target detection model are optimized using one or more techniques before being deployed to the airborne edge computing module, including: Model quantization: Converting the 32-bit floating-point numbers used to represent weights and activation values ​​in the model into low-precision 8-bit integers; Model pruning: Removing redundant parameters or structural units from a neural network; Knowledge distillation: Using a teacher model to guide the training of a student model, transferring knowledge into the student model.

[0013] As a preferred technical solution, the optimized low-light enhancement algorithm and target detection model are deployed on the target hardware platform, and the process is as follows: Export the trained and optimized model as an ONNX open standard intermediate format file; Use the inference optimizer of the target hardware platform to compile the ONNX file into an inference engine file optimized for the target hardware. During the compilation process, the inference optimizer performs optimization operations and deploys the optimized engine file to the onboard edge computing module, where the onboard application loads and executes the inference at runtime.

[0014] As a preferred technical solution, the inference optimizer performs optimization operations including: layer and tensor fusion, which merges multiple independent layers into a single CUDA kernel; precision calibration, which determines the optimal scaling factor by running on a calibration dataset when using INT8 quantization to minimize information loss; and kernel auto-tuning, which tests and selects the fastest algorithm implementation based on the target GPU model and data size.

[0015] As a second aspect of the present invention, a real-time image enhancement and exposure system for UAV inspection is provided. The system executes the real-time image enhancement and exposure method for UAV inspection as described above. The system logic architecture includes: a hardware layer and a software layer. The hardware layer includes: a drone platform, a visual sensor payload, an edge computing module, and a communication link; The software layer, running on top of the edge computing module, includes: The intelligent sensing and processing module further includes: Low-light enhancement submodule: Used for real-time enhancement of the original video stream; The target detection submodule is used to identify key targets from the enhanced video stream in real time and output the region of interest. Intelligent exposure control submodule: used to receive the region of interest, concentrate the metering weight of the automatic exposure algorithm on the region of interest, and generate an exposure control signal to adjust the exposure parameters of the visual sensor payload; The real-time video stream captured by the visual sensor payload is transmitted to the edge computing module; inside the edge computing module, the video stream is sequentially processed through the low-light enhancement submodule and the target detection submodule; the region of interest coordinates output by the target detection submodule are transmitted to the intelligent exposure control submodule; the exposure parameter control signal output by the intelligent exposure control submodule is transmitted to the visual sensor payload to dynamically adjust the camera's imaging parameters in real time.

[0016] As a preferred technical solution, the unmanned aerial vehicle platform adopts an industrial-grade multi-rotor unmanned aerial vehicle with RTK high positioning capability.

[0017] As a preferred technical solution, the visual sensor payload uses a visible light camera whose exposure parameters can be controlled in real time by an external signal.

[0018] As a preferred technical solution, the communication link supports 5G and / or 4G networks for the transmission of data and control signaling.

[0019] As a preferred technical solution, the dark light enhancement model running in the dark light enhancement submodule is a zero-reference learning model; the target detection model running in the target detection submodule is a single-stage detector or a hybrid detector based on Transformer; the dark light enhancement submodule and the target detection submodule are lightweighted before being deployed to the edge computing module, and are compiled and deployed on the edge computing module using an inference acceleration engine.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention proposes a real-time closed-loop feedback system for enhancement-recognition-exposure, ensuring optimal exposure for key targets at the imaging source. This resolves the vicious cycle and inherent contradiction of motion blur and high noise in nighttime inspections. Existing post-image processing techniques passively repair image information after it has been lost, with limited effectiveness; while this invention actively intervenes in the camera hardware before information loss, ensuring acquisition quality from the source and greatly improving the robustness and recognition accuracy of downstream models.

[0021] 2) This invention employs a zero-reference learning model (such as Zero-DCE) for low-light enhancement and an optimized lightweight detector (such as RT-DETR) for target recognition. This addresses the industry pain point of difficulty in obtaining paired training data and labeled data due to the special and dynamic nature of the scene. Simultaneously, the extreme lightweight nature and high inference speed of the selected model, combined with deployment optimization techniques such as TensorRT, ensures the feasibility of running complex AI algorithms in real-time on edge platforms with limited computing power. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the process for the real-time image enhancement and exposure method for nighttime inspection using a drone according to the present invention.

[0023] Figure 2 This is a diagram illustrating the overall architecture of the UAV nighttime inspection real-time image enhancement and exposure system of the present invention.

[0024] Figure 3 The images shown are comparisons of the image processing and exposure simulation results in this embodiment of the invention; where (a) is the original image under simulated low-light conditions; (b) is the intermediate result after end-side AI low-light enhancement; and (c) is the final image result after re-taking with intelligent exposure control.

[0025] Figure 4 This is a schematic diagram of the process for deploying the algorithm model on the client side in this invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] Example 1 The purpose of this invention is to overcome the problems of motion blur, high noise, and low signal-to-noise ratio caused by low-light environments in existing UAV night inspection technologies. These problems result in poor imaging quality and limited information acquisition capabilities of airborne vision systems. This invention aims to provide a method and system for generating high-quality, high-definition, and accurately exposed images in real time during flight through hardware and software co-optimization of airborne equipment, thereby providing high-quality data support for subsequent downstream defect identification tasks.

[0028] As a first aspect of the present invention, to address the difficulty of nighttime inspection imaging in scenarios such as rail transit and power line inspection, a real-time image enhancement and exposure method for drone nighttime inspection based on edge AI is provided, such as... Figure 1 As shown, the specific steps include: Step S1: Global Low-Light Enhancement. The real-time raw video stream captured by the UAV's onboard camera is acquired, and the onboard edge computing platform continuously and in real-time runs a lightweight low-light enhancement algorithm on the video stream to perform basic image preprocessing, thereby initially improving the overall image visibility and signal-to-noise ratio.

[0029] Step S2: Key Target Identification. On the video stream enhanced in Step S1, a lightweight target detection model is run to quickly identify and locate preset inspection targets (such as track fasteners, contact wire insulators, etc.), and accurately label and output the location coordinates and coverage area of ​​the target's region of interest.

[0030] Step S3: Intelligent Exposure of the Target Area. The coordinates of the region of interest identified in Step S2 are fed back to the UAV camera's control system in real time. The camera control system assigns the metering weight of its automatic exposure algorithm to the region of interest and dynamically adjusts exposure parameters (such as shutter speed, ISO sensitivity, gain, etc.) to ensure that the key inspection target receives optimal exposure in the next imaging, thereby capturing a clear image at the source of the image.

[0031] Specifically, step S3 involves motion-constrained intelligent exposure control based on region of interest (ROI) weighting. This step does not rely on the camera's general global automatic exposure algorithm, but instead employs an active parameter calculation strategy specifically for high-speed inspection scenarios, including the following sub-steps: S3-1: Constructing a dynamic photometric weighting mask (Weighting Mask Generation).

[0032] Based on the target ROI location coordinates output in step S2, a weight matrix consistent with the current video frame resolution is generated in real time in the image processing unit.

[0033] Set the pixel weights within the ROI region to high values ​​(e.g., ), set the weight of background pixels outside the ROI region to a very low value (e.g. Or 0).

[0034] The weighted average of the original image data is calculated using this weight matrix to obtain the weighted luminance value of the target area. This process eliminates the interference of complex backgrounds at night (such as bright streetlights or a pitch-black night sky) on the photometric system at the data level, ensuring that the photometric values ​​only reflect the brightness of the inspected target itself.

[0035] S3-2: Motion-Constrained Shutter Calculation based on flight kinematics.

[0036] To address the issue of motion blur during nighttime inspections, the system incorporates the current flight status data of the drone (provided by the onboard flight control system) to calculate the hard constraints of exposure parameters in real time.

[0037] Based on the drone's current ground flight speed V and ground sampling distance GSD Calculate the maximum permissible exposure time to ensure the image does not produce motion blur. .

[0038] The calculation formula is: in, k Allowable pixel blur threshold (e.g.) k= (0.5 pixels). This step ensures that the exposure time never exceeds the critical value that would cause image blur, which is something that conventional camera AE algorithms cannot achieve.

[0039] S3-3: Joint Parameter Solving & Compensation.

[0040] The exposure difference is calculated based on the weighted brightness value and the preset target brightness value, and the final parameters are calculated according to the priority strategy of "shutter limiting, gain compensation". Prioritize shutter speed: within the maximum permissible exposure time. Under the premise of increasing the shutter speed, brightness can be improved. Gain compensation: When the shutter speed reaches the maximum permissible exposure time If the image is still underexposed after the upper limit is reached, the system automatically calculates the required sensitivity gain (Analog / Digital Gain) and performs numerical compensation to forcibly increase the image brightness.

[0041] Closed-loop control: The calculated exact shutter speed and gain values ​​are written directly into the hardware registers through the camera control interface (API) to control the acquisition of the next frame image.

[0042] This invention aims to provide a method and system for generating high-quality, high-definition, and accurately exposed images in real time during flight through the coordinated optimization of onboard equipment's hardware and software, thereby providing high-quality data support for subsequent downstream defect identification tasks.

[0043] Specifically, the lightweight dark light enhancement algorithm in step S1 is preferably a zero-reference learning model.

[0044] Traditional deep learning-based low-light enhancement methods require supervised learning datasets of paired low-light and normal-light images, which are nearly impossible to obtain in the real-world dynamic scenarios of drone inspections. Unsupervised learning methods (such as GANs) are unstable during training and prone to artifacts.

[0045] This invention preferably employs a zero-reference learning model (e.g., the Zero-DCE model). This type of model requires no reference images during training; instead, it guides network learning by designing a set of task-related, reference-free loss functions (such as spatial consistency loss, exposure control loss, and color constancy loss). This approach fundamentally solves the core pain point of obtaining difficult-to-obtain training data in special scenarios such as rail transit.

[0046] In addition, the Zero-DCE model and its variants are extremely lightweight (e.g., the number of model parameters is only about 80,000) and can achieve extremely fast inference speeds (e.g., more than 100 FPS) on edge computing platforms, fully meeting the real-time requirements of UAV inspection.

[0047] In step S2, the lightweight target detection model is preferably a single-stage detector or a Transformer-based hybrid detector.

[0048] Traditional two-stage detectors, while highly accurate, are slow inference devices and cannot meet the real-time feedback required in step S3.

[0049] This invention preferably employs a high real-time detector, such as the YOLO series (e.g., YOLOv11) or the RT-DETR model. These models achieve an optimal balance between speed and accuracy (e.g., the RT-DETR-R50 can reach 108 FPS on a T4 GPU while achieving 53.1% mAP), ensuring sufficiently fast target recognition to provide millisecond-level real-time region of interest feedback to the camera exposure control system, thereby enabling effective closed-loop control.

[0050] In step S3, intelligent exposure of the target area is achieved through a hybrid intelligent exposure system based on the region of interest.

[0051] Traditional camera auto exposure algorithms typically meter the entire scene or the central area, which is prone to failure in high dynamic range scenes at night (such as bright traffic lights and dark track backgrounds).

[0052] After receiving the region of interest coordinates output in step S2, this invention allocates the vast majority (e.g., 80% to 100%) of the metering weights to pixels within the region of interest when calculating exposure parameters, while allocating very little (e.g., 0% to 20%) of the weights to pixels outside the region of interest. This mechanism ensures that exposure decisions are entirely biased towards key inspection targets, reducing interference from excessively dark or bright external areas and ensuring that object textures within the region of interest are clear and details are visible.

[0053] Furthermore, the method also includes step S4: algorithm lightweighting and edge deployment. Before deploying the algorithm model described in steps S1 and S2 to the edge computing module, it is lightweighted to meet the stringent requirements of limited computing power and power consumption of the airborne edge computing platform. Step S4 specifically includes: Step S4.1: Model Optimization. The model is optimized using one or more of the following techniques, including but not limited to: Model quantization: The 32-bit floating-point numbers used to represent weights and activation values ​​in the model are converted into low-precision 8-bit integers. This process can compress the model size by about 4 times and significantly improve computing speed and reduce energy consumption by leveraging the integer computing capabilities of edge computing hardware.

[0054] Model pruning: Removing redundant parameters or structural units from a neural network (e.g., structured pruning, removing entire convolutional kernel channels) to reduce model size and computational cost without significantly reducing accuracy.

[0055] Knowledge distillation: Using a large, high-precision "teacher model" to guide the training of a small, lightweight "student model" transfers "knowledge" into the student model, enabling it to achieve performance far exceeding what it could achieve on its own with a smaller model size.

[0056] Step S4.2: Model Compilation and Deployment. A standardized deployment process is used to efficiently deploy the optimized model on the target hardware platform. This process preferably includes: 1. Export models trained and optimized in frameworks such as PyTorch or TensorFlow to the ONNX open standard intermediate format.

[0057] 2. Using the inference optimizer (such as TensorRT SDK) of the target hardware platform (such as NVIDIA Jetson series), compile the ONNX file into an inference engine file that is highly optimized for the target hardware.

[0058] 3. During the compilation process, the inference optimizer performs key optimization operations, including: layer and tensor fusion (merging multiple independent layers into a single CUDA kernel to reduce kernel startup overhead and memory read / write operations), precision calibration (when using INT8 quantization, determining the optimal scaling factor by running on a calibration dataset to minimize information loss), and kernel auto-tuning (testing and selecting the fastest algorithm implementation based on the target GPU model and data size).

[0059] 4. Finally, the optimized engine file is deployed to the edge computing module, where it is loaded and executed by the onboard application at runtime.

[0060] Example 2 In a second aspect, this embodiment also provides a real-time image enhancement and exposure system for drone nighttime inspections based on edge AI, such as... Figure 2 As shown, the overall logical architecture of the system includes a hardware layer and a software layer.

[0061] The hardware layer includes: A drone platform: preferably an industrial-grade multi-rotor drone with high payload capacity, long endurance, and RTK high-precision positioning capability.

[0062] A visual sensor payload: preferably an integrated high-resolution, high-sensitivity visible light camera whose exposure parameters can be controlled in real time by an external signal.

[0063] An edge computing module, such as an NVIDIA Jetson Orin or Xavier series module, or a Qualcomm robotics platform, is used to perform high-intensity AI computing on the drone's edge.

[0064] A communication link: such as a 5G or 4G network, to ensure low-latency transmission of data and control signaling.

[0065] The software layer, running on top of the edge computing module, includes: An intelligent sensing and processing module, which is the core of this invention, further includes: A low-light enhancement submodule: used to perform step S1 in the method described in Example 1, to enhance the original video stream in real time.

[0066] A target detection submodule: used to perform step S2 in the method described in Example 1, to identify key targets in real time from the enhanced video stream and output the region of interest.

[0067] A smart exposure control submodule: used to perform step S3 in the method described in Embodiment 1, receive the region of interest and generate an exposure control signal to adjust the exposure parameters of the visual sensor payload.

[0068] The connection relationships between the modules in the system are as follows: The real-time video stream captured by the visual sensor payload is transmitted to the edge computing module; inside the edge computing module, the video stream is sequentially processed through the low-light enhancement submodule and the target detection submodule; the output of the target detection submodule (region of interest location coordinates) is transmitted to the intelligent exposure control submodule; the output of the intelligent exposure control submodule (exposure parameter control signal) is transmitted to the visual sensor payload to dynamically adjust the camera's imaging parameters in real time.

[0069] Example 3 As one specific implementation example of the present invention, this embodiment verifies the effectiveness of the image enhancement and exposure method based on edge AI proposed in the present invention in low-light inspection scenarios by constructing a set of simulation experimental systems.

[0070] 1. Experimental Environment and Configuration Image acquisition equipment: High-definition digital imaging equipment with manual exposure control is used to simulate the onboard camera of a drone.

[0071] Computing platform: A general-purpose embedded edge computing development board (integrated AI acceleration unit) is used to simulate an airborne edge computing module.

[0072] Algorithm model: Dark-light enhancement model: The Zero-DCE model is selected. This model has extremely low parameter count, making it suitable for operation on embedded devices with limited computing power.

[0073] Object detection model: A lightweight version of YOLOv11 was selected. This model maintains high accuracy while meeting the inference speed requirements of real-time video stream processing.

[0074] Model deployment steps are as follows Figure 4As shown: The above model was exported in ONNX format and compiled using an embedded platform-specific inference acceleration engine. INT8 quantization and operator fusion strategies were enabled during compilation. Experiments showed that on the embedded platform, the total time for dual-model serial inference was less than 30ms, meeting the real-time performance requirement of >30FPS. 2. Experimental process and results analysis as follows: Figure 3 As shown, the experiment simulated the extremely dark conditions of the track at night (illuminance <10 Lux) and observed key items.

[0075] Phase 1: Raw Imaging (Simulating Conventional Drone Photography) Figure 3 As shown in Figure a, when the intelligent function is not enabled, the device uses the default automatic exposure logic. Due to the extremely low ambient light, the main subject of the image is severely underexposed, making it almost impossible to discern the existence of the target object; furthermore, because the sensor automatically increases the ISO, the image is filled with a large amount of noise. Images in this state cannot be used for subsequent defect analysis.

[0076] Phase Two: Real-time Low-Light Enhancement. The original video stream is input into the embedded computing platform, and the Zero-DCE algorithm is run. For example... Figure 3 As shown in b, the image enhanced by the algorithm shows a significant improvement in overall brightness, making objects in dark areas clearly visible. Next, the object detection model successfully identified the target object in the enhanced image and output its region of interest (ROI) coordinates. This step verifies the necessity of the "enhancement before detection" strategy in extremely low-light conditions.

[0077] Phase 3: Intelligent Exposure Feedback. Based on the ROI coordinates output in step S2, the system calculates the brightness histogram of the area and determines the optimal combination of exposure parameters: reducing ISO to suppress noise, and appropriately adjusting the shutter speed to match the brightness requirements of the ROI. For example... Figure 3 As shown in Figure c, the image taken after applying the new parameters shows that the target object has better exposure and sharper texture details.

[0078] Experimental results show that the method proposed in this invention can effectively utilize the limited computing power on the edge and obtain high-quality, highly available inspection images in low-light environments at night through closed-loop linkage between the algorithm and imaging hardware.

[0079] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for real-time image enhancement and exposure for unmanned aerial vehicle inspection, characterized by the steps of include: The system acquires real-time raw video streams captured by the drone's onboard camera. The onboard edge computing platform runs a lightweight low-light enhancement algorithm to enhance the image of the raw video stream, improving the visibility and signal-to-noise ratio of the image. For the enhanced video stream, the airborne edge computing platform runs a lightweight target detection model to identify and locate the preset inspection targets, and marks and outputs the location coordinates and coverage area of ​​the target's region of interest. The location coordinates of the identified region of interest are fed back to the control system of the drone camera in real time. The camera control system then focuses the metering weight of its automatic exposure algorithm on the region of interest and dynamically adjusts the exposure parameters to capture images with clear details. 2.The unmanned aerial vehicle inspection real-time image enhancement and exposure method of claim 1, wherein, The dark light enhancement algorithm employs a zero-reference learning model, using a set of task-related, reference-free loss functions to guide network learning during the training process.

3. The method for real-time image enhancement and exposure during UAV inspection according to claim 1, characterized in that, The target detection model employs a single-stage detector or a hybrid detector based on Transformer.

4. The method for real-time image enhancement and exposure during UAV inspection according to claim 1, characterized in that, The metering weights of the automatic exposure algorithm are concentrated on the region of interest, specifically as follows: A dynamic metering weight mask is constructed, and a weight matrix with the same resolution as the current video frame is generated in real time based on the position coordinates of the target's region of interest. The pixel weights in the region of interest are set to high values, and the background pixel weights outside the region of interest are set to extremely low values. The weighted average of the original image data is calculated using a weight matrix to obtain the weighted brightness value.

5. The method for real-time image enhancement and exposure during UAV inspection according to claim 4, characterized in that, The specific details of the dynamic adjustment of exposure parameters are as follows: The maximum allowable exposure time of the shutter is calculated in real time, and the exposure difference is calculated based on the weighted brightness value of the target area and the preset target brightness value. Brightness can be increased by extending the shutter speed, provided that the maximum allowable exposure time is not exceeded. If the image is still underexposed after the shutter speed reaches the maximum allowable exposure time, the required ISO gain is calculated and numerical compensation is performed to increase the image brightness. The calculated shutter speed and ISO gain values ​​are written into the hardware register through the camera control interface to control the acquisition of the next frame image.

6. The method for real-time image enhancement and exposure during UAV inspection according to claim 1, characterized in that, The dark light enhancement algorithm and target detection model are optimized using one or more techniques before being deployed to the airborne edge computing module, including: Model quantization: Converting the 32-bit floating-point numbers used to represent weights and activation values ​​in the model into low-precision 8-bit integers; Model pruning: Removing redundant parameters or structural units from a neural network; Knowledge distillation: Using a teacher model to guide the training of a student model, transferring knowledge into the student model.

7. The method for real-time image enhancement and exposure during UAV inspection according to claim 1, characterized in that, The optimized low-light enhancement algorithm and object detection model are deployed on the target hardware platform, and the process is as follows: Export the trained and optimized model as an ONNX open standard intermediate format file; Use the inference optimizer of the target hardware platform to compile the ONNX file into an inference engine file optimized for the target hardware. During the compilation process, the inference optimizer performs optimization operations and deploys the optimized engine file to the onboard edge computing module, where the onboard application loads and executes the inference at runtime.

8. The method for real-time image enhancement and exposure during UAV inspection according to claim 7, characterized in that, The inference optimizer performs optimization operations including: layer and tensor fusion, which merges multiple independent layers into a single CUDA kernel; precision calibration, which determines the optimal scaling factor by running on a calibration dataset to minimize information loss when using INT8 quantization; and kernel auto-tuning, which tests and selects the fastest algorithm implementation based on the target GPU model and data size.

9. A real-time image enhancement and exposure system for unmanned aerial vehicle (UAV) inspection, characterized in that, The system executes the real-time image enhancement and exposure method for UAV inspection as described in any one of claims 1-8, and the system logic architecture includes: a hardware layer and a software layer; The hardware layer includes: a drone platform, a visual sensor payload, an edge computing module, and a communication link; The software layer, running on top of the edge computing module, includes: The intelligent sensing and processing module further includes: Low-light enhancement submodule: Used for real-time enhancement of the original video stream; The target detection submodule is used to identify key targets from the enhanced video stream in real time and output the region of interest. Intelligent exposure control submodule: used to receive the region of interest, concentrate the metering weight of the automatic exposure algorithm on the region of interest, and generate an exposure control signal to adjust the exposure parameters of the visual sensor payload; The real-time video stream captured by the visual sensor payload is transmitted to the edge computing module; inside the edge computing module, the video stream is sequentially processed through the low-light enhancement submodule and the target detection submodule; the region of interest coordinates output by the target detection submodule are transmitted to the intelligent exposure control submodule; the exposure parameter control signal output by the intelligent exposure control submodule is transmitted to the visual sensor payload to dynamically adjust the camera's imaging parameters in real time.

10. A real-time image enhancement and exposure system for UAV inspection according to claim 9, characterized in that, The unmanned aerial vehicle platform adopts an industrial-grade multi-rotor drone with RTK high positioning capability.

11. A real-time image enhancement and exposure system for UAV inspection according to claim 9, characterized in that, The visual sensor payload is a visible light camera whose exposure parameters can be controlled in real time by an external signal.

12. The real-time image enhancement and exposure system for UAV inspection according to claim 9, characterized in that, The communication link supports 5G and / or 4G networks for the transmission of data and control signaling.

13. The real-time image enhancement and exposure system for UAV inspection according to claim 9, characterized in that, The dark light enhancement model running in the dark light enhancement submodule uses a zero-reference learning model; The target detection model running in the target detection submodule is either a single-stage detector or a hybrid detector based on Transformer. The dark light enhancement submodule and the target detection submodule are lightweighted before being deployed to the edge computing module, and are compiled and deployed on the edge computing module using an inference acceleration engine.

Images