A polarization light-suppressed self-cleaning high-quality inspection imaging generation method

Abstract

The present application relates to the technical field of image processing, and particularly relates to a polarization light-suppressed self-cleaning high-quality inspection imaging generation method, a preview image is collected and partitioned, a light spot detection model is used to evaluate the number of light spots in each region, and a gimbal is controlled to adjust to the region with the least light spots; a polarizing mirror is rotated and light spot intensity is monitored in real time, and the angle at which the intensity is the least is locked; an image is collected and a definition index is calculated, and if the definition index is below a first threshold, first-level cleaning is triggered, and if the definition index does not recover to a second threshold within a preset time, second-level cleaning is triggered; the present application combines gimbal coarse adjustment and polarizing mirror fine light suppression, avoids frequent adjustment of the attitude of a UAV, reduces the calculation burden and flight risk, and at the same time, based on the feedback of the definition index, the lens is maintained on demand through hierarchical self-cleaning, dynamic reflected light and lens contamination are effectively suppressed, and a high-quality inspection image is generated.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for generating high-quality inspection images with polarization suppression and self-cleaning properties. Background Technology

[0002] When using drones to inspect offshore wind turbine blades, the imaging quality is constrained by two core factors: (1) Strong dynamic reflection light interference: The sea surface forms a dynamically changing mirror due to the undulation of waves, which reflects sunlight onto the wind turbine blades, forming "moving light spots" with rapidly changing position, shape and intensity. These light spots cause local overexposure in the image, completely covering key defects such as cracks and corrosion.

[0003] (2) Lens contamination interference: The high salt fog environment at sea is prone to condensation on the optical lens of the UAV, forming water droplets or salt crystals, resulting in overall image blurring and reduced contrast.

[0004] Currently, the mainstream solution is "active light avoidance," which uses computer vision to identify light spots and dynamically adjusts the drone's attitude to avoid them. This method is effective in static scenarios such as solar panels, but it has serious limitations when applied to offshore wind turbines. The high-speed rotation of the turbine blades, combined with the irregular changes in the direction of the light source caused by sea waves, results in the light spots moving rapidly and randomly in the image. Relying on attitude adjustments for tracking and avoidance would lead to violent and frequent maneuvers by the drone. This not only greatly increases flight risks in strong winds and causes motion blur, but also severely sacrifices inspection efficiency due to computational complexity and action delays. Furthermore, making large-scale attitude adjustments to avoid tiny light spots is extremely inefficient.

[0005] It is evident that the existing technology has the following drawbacks: (1) Poor dynamic adaptability: The "active light avoidance method" based on attitude adjustment cannot effectively track the dynamic light spot with irregular movement at sea. The real-time requirements of the algorithm conflict with the performance limitations of the airborne edge computing equipment.

[0006] (2) The contradiction between stability and efficiency: Frequent and drastic attitude adjustments sacrifice flight stability and shooting window, resulting in blurred images, inefficient inspection paths, and increased energy consumption.

[0007] Therefore, there is an urgent need for a new inspection image generation scheme that can minimize the interference of strong dynamic reflected light and lens contamination during imaging, while ensuring equipment adaptability and the stability and efficiency of image generation. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a self-cleaning, high-quality inspection imaging generation method with polarization suppression, which can minimize the interference of strong dynamic reflected light and lens contamination during imaging, while ensuring equipment adaptability and the stability and efficiency of image generation.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for generating high-quality, self-cleaning inspection images with polarization suppression includes the following steps: S1. Acquire a preview image of the inspection target, divide the preview image into regions, use a pre-trained spot detection model to evaluate the number of spots in each region, and control the camera pan-tilt unit to adjust towards the region with the fewest spots. S2. Control the motorized polarizer to rotate around the optical axis and monitor the light spot intensity in real time, and lock the motorized polarizer to the angle when the light spot intensity is minimum; S3. Acquire images of the inspection target and calculate the image clarity index in real time. When the clarity index is lower than the first preset threshold, trigger the first-level cleaning action. If the clarity index does not recover to the second preset threshold within a preset time after the first-level cleaning, trigger the second-level cleaning action.

[0010] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A storage medium storing a computer program thereon, which, when executed, implements the steps in the above-described method for generating a self-cleaning, high-quality inspection image with polarization suppression.

[0011] The beneficial effects of this invention are as follows: This application provides a self-cleaning, high-quality inspection imaging generation method for polarization-suppressed light. It employs regional spot evaluation and gimbal fine-tuning, requiring only a one-time small-angle adjustment of the gimbal to shift the field of view to the area with the fewest light spots. This avoids frequent and significant adjustments to the UAV's attitude to track irregularly moving dynamic light spots, thus significantly reducing flight risks and computational burden. By optimizing and locking the optimal polarization angle through polarizing mirror rotation, it statically filters out reflected light from specific directions at the physical optics level. Once locked, it is unaffected by light spot movement, compensating for the shortcomings of real-time tracking in dynamic scenes. Furthermore, it introduces a graded self-cleaning mechanism based on image sharpness, treating lens contamination as an independent but coupled factor in a closed-loop manner. This overcomes the shortcomings of existing technologies that address single problems in isolation, ensuring that the light suppression parameters remain effective under clean optical front-ends. This invention, in a lightweight, low-latency, and closed-loop controllable manner, simultaneously suppresses dynamic reflected light and lens contamination, significantly improving the robustness and imaging quality of dynamic maritime inspections. Attached Figure Description

[0012] Figure 1This is a flowchart of a self-cleaning, high-quality inspection imaging generation method with polarization suppression, according to an embodiment of the present invention. Detailed Implementation

[0013] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0014] The self-cleaning, high-quality inspection image generation method with polarization suppression described above is applicable to UAV inspection scenarios involving strong reflected light and lens contamination risks, such as offshore wind turbine blades, photovoltaic power stations, and bridges. The following detailed implementation method illustrates this approach: Please refer to Figure 1 One embodiment of the present invention is as follows: A method for generating high-quality, self-cleaning inspection images with polarization suppression includes the following steps: S1. Acquire a preview image of the inspection target, divide the preview image into regions, use a pre-trained spot detection model to evaluate the number of spots in each region, and control the camera pan-tilt unit to adjust towards the region with the fewest spots. The spot detection model was trained using the YOLOv11 model. The pre-trained spot detection model is used to evaluate the number of spots in each region as follows: The images of each region are sequentially input into the spot detection model, and the number of spots in each region is determined based on the detection results of the spot detection model.

[0015] In this embodiment, after the UAV hovers at the inspection starting point, it activates the gimbal camera to acquire a global preview image. The preview image is divided into N equal-area regions (N=4 in this embodiment). The image of each region is sequentially input into a lightweight spot detection model (e.g., YOLOv11) deployed on an airborne edge computing device. The model outputs the total number of spots detected in each region. The number of spots in the four regions is compared, and the direction of the region with the fewest spots is selected. The gimbal is then fine-tuned in that direction by a fixed small angle (e.g., 2° to 5°). The purpose of this step is to shift the main field of view to a relatively "clean" background through a one-time coarse adjustment, significantly reducing the load on subsequent polarizer optimization and avoiding continuous shaking caused by tracking fine spots.

[0016] Specifically, the adjustment steps in this embodiment are exemplified as follows: For example, if it is the lower right corner area, rotate the horizontal angle of the gimbal 2 degrees to the right. If the number of light spots does not decrease, increase it by one degree, up to a maximum of five degrees. If the maximum is reached and the number of light spots still does not decrease, adjust the vertical angle on the original horizontal basis. The adjustment range is 1-5 degrees, until the number of light spots is reduced to 50% of the original.

[0017] S2. Control the motorized polarizer to rotate around the optical axis and monitor the light spot intensity in real time, and lock the motorized polarizer to the angle at which the light spot intensity is minimal.

[0018] In this private setup, keeping the gimbal angle constant, the motorized polarizing filter is rotated uniformly around its optical axis for one full rotation (360°). During rotation, the overall light spot intensity in the camera image is monitored in real time, for example, by calculating the sum of bright pixels or the area-weighted value of the light spot region. The rotation angle θ_min of the polarizing filter corresponding to the point where the light spot intensity reaches its minimum value is recorded. After one rotation, the polarizing filter is controlled to rotate in the opposite direction to the θ_min angle and locked. This step, by changing the polarization state, physically filters out most of the reflected light from specific directions, achieving "static" suppression of residual strong reflected light, unaffected by subsequent minor movements of the light spot.

[0019] S3. Acquire images of the inspection target and calculate the image clarity index in real time. When the clarity index is lower than the first preset threshold, trigger the first-level cleaning action. If the clarity index does not recover to the second preset threshold within a preset time after the first-level cleaning, trigger the second-level cleaning action.

[0020] In this embodiment, after the light suppression state is locked, inspection images are continuously acquired, and the sharpness index D of each frame is calculated in real time. In this embodiment, the sharpness index is calculated as follows: Convert the image to grayscale and divide it into several non-overlapping image blocks; In this embodiment, the acquired color image is converted into a grayscale image. Since the goal is to suppress light to the maximum extent, some light spots may still be present in extreme environments. To avoid the remaining light spots interfering with the statistics, the preprocessed image is first divided into 16 x 16 non-overlapping blocks.

[0021] Calculate the average brightness and brightness variance of each small block. Based on the average brightness, brightness variance, and a preset threshold, remove abnormal image blocks to obtain valid image blocks. In this embodiment, the average brightness (Mean) and brightness variance (Var) of each small block are calculated. A threshold is set to remove blocks that may have excessively high Mean due to extreme light spots, or blocks with low Var due to a lack of texture, such as those representing a uniform sky or sea surface (with fine texture like a wind turbine). Only valid image blocks with rich textures and unaffected by extreme light spots are retained.

[0022] For each of the effective image blocks, the Sobel operator is used to calculate the gradient magnitude map in the x and y directions, and based on each of the gradient magnitude maps, the image sharpness index is calculated; Based on each of the aforementioned gradient magnitude maps, the image sharpness index is calculated as follows: Based on each of the gradient magnitude maps, the image block score of each of the effective image blocks is calculated; The image block fractions are expressed as the sum of squared gradients or the sum of gradient magnitudes; The sharpness index is obtained by calculating the average or median of the image block scores of all the valid image blocks.

[0023] In this embodiment, the Sobel operator is used to calculate the gradient magnitude map G(x, y) in the x and y directions for each valid image block, and then the sum of squared gradients (or the sum of gradient magnitudes) of all pixels in each block is calculated as the image block score Block_Score for that block.

[0024] Sum of squared gradients: Block_Score = ΣΣG(x, y)²; Gradient magnitude sum: Block_Score = ΣΣ|G(x, y)|; The final sharpness index D is obtained by averaging (or medianing) the Block_Scores of all valid image blocks.

[0025] The primary cleaning action includes starting the centrifugal spin dryer and running it at a first speed; The second cleaning action includes controlling the centrifugal spin dryer to run at a second speed and simultaneously starting the cleaning agent spraying unit for compound cleaning; The second speed is greater than the first speed.

[0026] Step S3 also includes the following steps: If the clarity index fails to recover to the second preset threshold within a preset time after the completion of the secondary cleaning action, the self-cleaning is deemed to have failed, and an alarm and return-to-base suggestion are sent to the ground station.

[0027] In this embodiment, a first preset threshold Th_low (e.g., 0.3) and a second preset threshold Th_high (e.g., 0.6) are set. When D < Th_low, a first-level cleaning action is triggered: the centrifugal drying device is started and runs at a first speed V (e.g., 3000 rpm) to attempt to remove water droplets or mist from the mirror surface. If, after the first-level cleaning is started, D does not recover to above Th_high within T seconds (e.g., 5 seconds), it is determined to be heavily contaminated (e.g., oil stains, salt crystals), and a second-level cleaning action is triggered: the centrifugal speed is increased to a second speed 2V (e.g., 6000 rpm), and the alcohol spraying unit is simultaneously started to spray a small amount of alcohol onto the mirror surface for compound cleaning. If, after the second-level cleaning is completed, D still cannot recover to Th_high within a preset time (e.g., 10 seconds), the system determines that the self-cleaning has failed, automatically sends an alarm to the ground station and suggests returning to base to prevent invalid inspections.

[0028] S4. Control the motorized polarizer to rotate around the optical axis again, and monitor the light spot intensity in real time. Lock the motorized polarizer to the angle when the light spot intensity is minimum, and calibrate the polarization state change caused by cleaning.

[0029] In this embodiment, after the self-cleaning operation (level 1 or level 2 cleaning) is completed and the clarity is restored to Th_high or above, since the cleaning process may change the surface state of the lens or a trace amount of residual liquid may affect the polarization characteristics, the system automatically re-executes the polarization mirror rotation optimization: it controls the motorized polarization mirror to rotate one revolution again, searches for the polarization angle with the minimum current light spot intensity and locks it, so as to calibrate the change in polarization state and ensure that the light suppression effect continues to be optimal.

[0030] Embodiment 2 of the present invention is as follows: A self-cleaning, high-quality inspection imaging generation method with polarization suppression is proposed. The difference between this method and the first embodiment is that the spot detection model in this embodiment adopts SDPNet (Spot Detection and Polarization-aware Network). This scheme uses YOLOv11 as a baseline and improves detection accuracy and inference speed through lightweight structure, dynamic feature enhancement, dedicated loss function and data augmentation strategy.

[0031] (1) Model structure: The overall architecture adopts a single-stage anchor-free architecture, with core improvements including: Lightweight backbone network (Efficient-MobileNetV4): The CSP layer in YOLOv11 is replaced by an inverted residual module combined with a dynamic activation function (DY-ReLU).

[0032] Separable self-attention is introduced in stages 3 and 4 to enhance contextual modeling of irregularly moving light spots while maintaining low FLOPs.

[0033] The output consists of three feature layers (P3, P4, P5), corresponding to small, medium, and large light spots, respectively.

[0034] Dynamic Receptive Field Enhancement Module (DRFEM): Located in the Neck layer, it performs deformable convolution (Deformable Conv v3) on multi-scale features, enabling the sampling points to adapt to the geometric deformation of the light spot (the stretching and distortion of the light spot caused by wave reflection).

[0035] Combined with lightweight channel attention (ECA-Net), it suppresses background noise (such as leaf texture and sea ripples).

[0036] Detection head improvement (Polarization-Aware Head): In addition to the standard bounding box regression and classification, a polarization state regression branch is added: predicting the optimal suppressed polarization angle (a continuous value from 0 to 180°) corresponding to the current spot, providing prior information for subsequent polarization mirror control. This branch is only used for knowledge distillation during the training phase and may not be output during inference.

[0037] The classification branch outputs "spot confidence level", and the regression branch outputs bounding box offset.

[0038] The overall parameter scale is controlled within 3.5M, and can reach >120FPS on Jetson Orin NX.

[0039] (2) Loss function design: A joint loss function is used, which consists of three parts: ; in, L cls This represents the classification loss, which measures the accuracy of predicting the category of the light spot (i.e., determining whether an image patch is a light spot). L box This represents the bounding box regression loss, which measures the difference between the predicted spot location and the actual spot location. λ pol The weighting coefficient representing the polarization angle-assisted loss is used to balance the importance of the three types of losses, and is usually taken as 0.1 to 0.5. L pol This represents the polarization angle-assisted loss, used only in the teacher model or distillation training phase, to learn the optimal suppressed polarization angle of the light spot.

[0040] Classification loss L cls Varifocal Loss is used to handle the extreme imbalance between positive and negative samples (the proportion of spot pixels is usually <5%).

[0041] ; in, N pos This represents the total number of positive samples (real light spots), used to normalize the loss and avoid an imbalance between the number of positive and negative samples. This indicates the indicator function when the prediction is a positive sample ( p i If the value is greater than 0, set the value to 1; otherwise, set the value to 0. Variational Loss only focuses on positive samples, reducing the interference from negative samples. p i This represents the probability predicted by the model that the location belongs to the spot category (after sigmoid normalization, ranging from 0 to 1).q i The target quality score is usually the center value of the IoU between the predicted bounding box and the ground truth bounding box (e.g., the weighted sum of IoU and the distance to the center point). The higher the value, the better the quality of the positive sample.

[0042] Bounding box regression loss L box Using CIoU Loss+ pixel-level gradient guidance: Since the edges of light spots are usually blurry, an edge-aware term is introduced: the variance of the gradient magnitude within the predicted box is calculated. If the variance is small (the texture inside the light spot is uniform), the regression penalty is reduced to avoid overfitting noise.

[0043] ; in, CIoU α represents Complete IoU, which introduces center point distance and aspect ratio consistency on the basis of IoU, and takes a value range of [-1, 1]. The larger the value, the higher the overlap between the predicted box and the ground truth box; α represents the weight coefficient of the edge perception term, which controls the correction strength of the gradient variance term to the regression loss, and is usually taken as 0.1 to 0.3. This represents the standard deviation of the image gradient magnitude within the prediction box region.

[0044] Polarization angle-assisted loss L pol Used for distilling the teacher model: ; The true polarization angle is obtained through offline calibration (e.g., taking pictures of the same light spot at different polarizing filter angles and then calculating its optimal suppression angle).

[0045] in, This represents the optimal suppression polarization angle corresponding to the i-th spot predicted by the model, ranging from 0° to 180°; This represents the true optimal suppression polarization angle calibrated offline, obtained through experiments (e.g., selecting the darkest angle of the light spot after taking a picture with a rotating polarizing filter). The indicator function is defined as follows: the intersection-union ratio (IoU) of the i-th predicted box and the ground truth box is (i=1 to 1). IoU i The polarization angle loss of the light spot is only included when the value is greater than 0.5. This ensures that the polarization angle is constrained only when the detection is correct, thus avoiding random regression.

[0046] (3) Training data source: Real-world data collection: Using a drone equipped with a motorized polarizing filter, at least 5000 images were captured near offshore wind turbines at different polarization angles (e.g., 0°, 45°, 90°, 135°). The bounding boxes of the light spots and the corresponding optimal suppression polarization angles were manually annotated. A hybrid "polygon + rectangle" format was used for annotation to accommodate irregular light spots.

[0047] Simulation generation: Using Unreal Engine 4, an ocean-wind turbine-light dynamics model was built to simulate the light spot morphology under different solar altitude angles and wave levels, and annotated synthetic images were automatically generated.

[0048] In addition, data augmentation strategies need to be implemented on the data: Elastic deformation (simulating waves) is applied to the light spot region in the existing image, and the light spot is randomly translated and scaled to generate motion blur between sequential frames.

[0049] Images taken at different polarization angles in the same scene are weighted and fused to generate a "virtual polarization state" image that lies between the two, thus broadening the model's ability to generalize to polarization changes.

[0050] Randomly add Gaussian noise, defocus blur, and fog effects to simulate lens contamination or low-light conditions.

[0051] For images with sparse light spots, the number of positive samples is increased by copying and pasting the light spots (while preserving the perspective transformation).

[0052] Split the dataset: Training set: 70% (including 80% real data and 20% synthetic data); Validation set: 15% (all true); Test set: 15% (all real and from different flight missions).

[0053] (4) Training strategy: Optimizer: AdamW, initial learning rate 0.001, cosine annealing decay.

[0054] Batch size: 32 (for a single RTX 4090).

[0055] Distillation training: First, train the teacher model on synthetic data using a large model (such as YOLOv9-E), and then perform knowledge distillation on SDPNet (focusing on feature map alignment and polarization angle regression).

[0056] Training cycle: 300 epochs, early stopping mechanism (if the validation set mAP@0.5 does not improve for 20 consecutive epochs, the training will stop).

[0057] Embodiment 3 of the present invention is as follows: A storage medium storing a computer program, which, when executed, implements the steps in the polarization suppression self-cleaning high-quality inspection imaging generation method described in Embodiment 1 or 2 above.

[0058] This invention presents a self-cleaning, high-quality inspection imaging generation method with polarization suppression. Through a two-stage strategy of coarse gimbal adjustment and fine polarization mirror suppression, it effectively suppresses dynamic light spots without relying on high-frequency attitude adjustments, resulting in low computational burden and making it particularly suitable for edge device deployments. Simultaneously, it introduces a hierarchical self-cleaning closed loop based on sharpness feedback, enabling on-demand maintenance and ensuring the cleanliness of the optical front end. Ultimately, the entire method improves the robustness and efficiency of inspection imaging in dynamic maritime environments.

[0059] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for generating high-quality, self-cleaning inspection images with polarization suppression, characterized in that, Including the following steps: S1. Acquire a preview image of the inspection target, divide the preview image into regions, use a pre-trained spot detection model to evaluate the number of spots in each region, and control the camera pan-tilt unit to adjust towards the region with the fewest spots. S2. Control the motorized polarizer to rotate around the optical axis and monitor the light spot intensity in real time, and lock the motorized polarizer to the angle when the light spot intensity is minimum; S3. Acquire images of the inspection target and calculate the image clarity index in real time. When the clarity index is lower than the first preset threshold, trigger the first-level cleaning action. If the clarity index does not recover to the second preset threshold within a preset time after the first-level cleaning, trigger the second-level cleaning action.

2. The method for generating high-quality inspection images with polarization suppression according to claim 1, characterized in that, The primary cleaning action includes starting the centrifugal spin dryer and running it at a first speed; The second cleaning action includes controlling the centrifugal spin dryer to run at a second speed and simultaneously starting the cleaning agent spraying unit for compound cleaning; The second speed is greater than the first speed.

3. The method for generating high-quality inspection images with polarization suppression according to claim 1, characterized in that, Step S3 also includes the following steps: If the clarity index fails to recover to the second preset threshold within a preset time after the completion of the secondary cleaning action, the self-cleaning is deemed to have failed, and an alarm and return-to-base suggestion are sent to the ground station.

4. The method for generating high-quality inspection images with polarization suppression according to claim 1, characterized in that, It also includes the following steps: S4. Control the motorized polarizer to rotate around the optical axis again, and monitor the light spot intensity in real time. Lock the motorized polarizer to the angle when the light spot intensity is minimum, and calibrate the polarization state change caused by cleaning.

5. The method for generating high-quality inspection images with polarization suppression according to claim 1, characterized in that, The calculation of the sharpness index includes the following steps: Convert the image to grayscale and divide it into several non-overlapping image blocks; Calculate the average brightness and brightness variance of each small block. Based on the average brightness, brightness variance, and a preset threshold, remove abnormal image blocks to obtain valid image blocks. For each valid image block, the Sobel operator is used to calculate the gradient magnitude map in the x and y directions, and based on each gradient magnitude map, the image sharpness index is calculated.

6. The method for generating high-quality inspection images with polarization suppression according to claim 5, characterized in that, Based on each of the aforementioned gradient magnitude maps, the image sharpness index is calculated as follows: Based on each of the gradient magnitude maps, the image block score of each of the effective image blocks is calculated; The image block fractions are expressed as the sum of squared gradients or the sum of gradient magnitudes; The sharpness index is obtained by calculating the average or median of the image block scores of all the valid image blocks.

7. The method for generating high-quality inspection images with polarization suppression according to claim 1, characterized in that, The spot detection model was trained using the YOLOv11 model.

8. The method for generating high-quality inspection images with polarization suppression according to claim 1, characterized in that, The pre-trained spot detection model is used to evaluate the number of spots in each region as follows: The images of each region are sequentially input into the spot detection model, and the number of spots in each region is determined based on the detection results of the spot detection model.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the steps in the self-cleaning high-quality inspection imaging generation method for polarization suppression as described in any one of claims 1-8.

Images