Novel unmanned aerial vehicle aerial photo photovoltaic defect data generation method

Abstract

The invention discloses a novel unmanned aerial vehicle aerial photo photovoltaic defect data generation method, and belongs to the technical field of photovoltaic power generation, and the method comprises the following steps: 1, an unmanned aerial vehicle carries a camera, carries out the full-width scanning of a photovoltaic panel according to a preset route, obtains an aerial photo Iraw, and an attitude sensor obtains a pitch angle theta in real time; 2, correcting the aerial image Iraw for the flight height h and the pitch angle theta to obtain an aerial image Iraw after geometric correction; step 3, taking defect data of the current solar cell panel or defects such as cracks, shielding and the like shot by an industrial camera at a distance of 1m as a close-shot image Iear, and fusing the close-shot image Iear and the image Iwarp after geometric correction to obtain a fused chromatic value Cout; according to the method, the highly adaptive LAB color space hybrid model is established, the close-shot defect chromaticity and the aerial photography background texture are dynamically fused, artifacts caused by color gamut differences between devices are eliminated, the problem of color gamut deviation of aerial photography images and close-shot defect data is solved, and the generated images are superior to those generated by a traditional RGB fusion mode.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and specifically to a novel method for generating photovoltaic defect data from drone aerial photography. Background Technology

[0002] With the rapid development of photovoltaic power generation technology, photovoltaic module inspection technology based on drone platforms has been widely used due to its significant advantages such as high efficiency and safety. By equipping high-resolution camera devices and combining them with computer vision technology, intelligent inspection systems can be built to achieve automated defect detection of photovoltaic modules. However, current technology applications face a key bottleneck: the scarcity of photovoltaic module defect samples, especially the lack of data on typical faults such as scratches and damage, makes it difficult to meet the needs of deep learning models for large-scale, high-quality training data, severely restricting the improvement of defect recognition accuracy.

[0003] The existing technology has the following drawbacks:

[0004] Data generation lacks physical authenticity: Traditional methods directly integrate near-field defects with aerial backgrounds, ignoring the coupling effect of flight altitude on color, geometry, and noise features, resulting in significant deviations between the generated data and the actual aerial distribution (e.g., LAB color difference ΔE>15).

[0005] Insufficient color gamut offset correction capability: Existing color transfer methods (such as histogram matching) have not established a quantitative mapping relationship between height and chromaticity, and cannot adaptively adjust the mixing weight within the height range of 10-100 meters, resulting in color block separation in the fused image.

[0006] Incomplete geometric distortion compensation: The affine transformation model with fixed parameters cannot adapt to changes in flight altitude, and the scale compression effect of high-altitude images (h>50m) leads to distortion of defect size.

[0007] The illumination modeling is too simplistic: most methods only use a single illumination pattern to generate data, without considering the differences in illumination intensity and scattering at different times (noon, cloudy, dusk), resulting in poor generalization ability of the model under complex illumination.

[0008] Based on this, the present invention designs a novel method for generating photovoltaic defect data by drone aerial photography to adapt to complex and ever-changing dynamic environments and meet practical application needs. Summary of the Invention

[0009] To address the aforementioned shortcomings of existing technologies, this invention provides a novel method for generating photovoltaic defect data from drone aerial photography.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A novel method for generating photovoltaic defect data from drone aerial photography includes the following steps:

[0012] Step 1: The drone, equipped with a camera, performs a full-frame scan of the photovoltaic panels along a preset flight path to obtain aerial images (I). raw The attitude sensor acquires the pitch angle θ in real time;

[0013] Step 2: Compare the aerial image I with the flight altitude h and pitch angle θ. raw Correction is performed to obtain the geometrically corrected aerial image I. raw ;

[0014] Step 3: Use the current solar panel defect data or take close-up images of defects such as cracks and shading at a distance of 1m using an industrial camera. near Close-up image I near Image I after geometric correction warped The fusion yields the fused chromaticity value C. out With brightness channel L aerial ;

[0015] Step 4: Combine the blended chromaticity values ​​C out With brightness channel L aerial The fused illumination parameters are added to the aerial image brightness channel to obtain a fused image with degraded illumination uniformity. degraded Specifically, it adopts three types of lighting modes;

[0016] Step 5: Process the degraded image I degraded Perform geometric enhancement and spectral enhancement;

[0017] Step 6: Use the pre-trained Mask R-CNN model to initially label defective regions in the geometrically enhanced and spectrally enhanced images, and then manually correct any missing or incorrectly labeled regions.

[0018] Furthermore, the specific steps of step 2 are as follows:

[0019] Step 21, using the original aerial image I raw The flight altitude h determines the deformation scaling factor s;

[0020] The deformation scaling factor s is calculated as follows:

[0021] s = h / 50, (h∈[10,100]m);

[0022] Step 22, Pitch angle θ, Original aerial image I raw And the deformation scaling factor s is embedded in the affine transformation matrix M, for the original aerial image I raw By applying affine transformation, we obtain the geometrically corrected image I. warped The specific calculations are as follows:

[0023] I warped =warpAffine(I raw ,M);

[0024]

[0025] Furthermore, the specific steps for step 3 are as follows:

[0026] Step 31, Near-field image I near Aerial images after geometric correction I warped Convert from RGB color space to LAB color space, and separate the luminance (L) and chroma (A / B) channels to obtain the luminance channel L. aerial C, color value of close-up image near and the background chromaticity value C of drone aerial photography aerial ;

[0027] Step 32: Based on the hybrid weighting, adjust the chromaticity values ​​C of the close-up image and the aerial background. aerial Weighted fusion is performed to obtain the fused chromaticity value C. out .

[0028] Furthermore, C out The specific calculations are as follows;

[0029] C out =α·C near +(1-α)·C aerial

[0030] α = 0.9 - 0.2 (h / 10⁻¹);

[0031] α is the blending weight of the LAB channel.

[0032] Furthermore, step 4 specifically adopts three types of lighting modes.

[0033] Furthermore, the specific steps for step 4 are as follows:

[0034] Step 41: Define the parameters for the three types of lighting modes;

[0035] Step 42: Set the brightness channel L aerial By integrating the lighting characteristics, the adjusted aerial brightness L′ is obtained, and the specific calculation is as follows:

[0036] L′=β·L aerial +N(0,σ 2 );

[0037] N(0,σ 2 ) represents the Gaussian noise term, β is the brightness gain adjustment, and σ is the degree of dispersion of the light intensity distribution;

[0038] Step 43: Merge LAB images (L′, C out The degraded image I was obtained. degraded The specific calculations are as follows:

[0039] I degraded =Blur(LAB2RGB(L′,C out ),k(h));

[0040] Where k(h)) represents the fuzzy kernel, which is set to 0.5h.

[0041] Furthermore, for the noon mode: the specific parameters for the three lighting modes are as follows:

[0042] HSV brightness gain coefficient β = 1.2, variance σ = 0.05;

[0043] Cloudy day pattern: β = 0.4, variance σ = 0.15;

[0044] Dusk mode: β = 0.8, variance σ = 0.1.

[0045] Furthermore, geometric enhancements include random rotation, horizontal / vertical rotation, and random cropping;

[0046] Spectral enhancement includes random adjustments to saturation and hue.

[0047] This invention establishes a highly adaptive LAB color space mixing model to dynamically fuse the chromaticity of near-field defects with the texture of aerial backgrounds, eliminating artifacts caused by color gamut differences between devices and solving the color gamut offset problem between aerial images and near-field defect data. The resulting image is superior to the traditional RGB fusion method.

[0048] This invention proposes a dynamic geometric deformation algorithm based on flight altitude and pitch angle. It corrects image distortion through an affine transformation matrix and introduces an altitude-related scaling factor (s = h / 50) to compensate for scale compression, thereby compensating for the geometric distortion and scale compression effects of high-altitude imaging.

[0049] This invention constructs a multi-time-period illumination model (noon, cloudy, dusk), and combines HSV brightness channel adjustment and Poisson fusion algorithm to simulate the light intensity distribution and scattering effect of real scenes, thereby improving the physical realism and environmental adaptability of defect data. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0051] Figure 1 This is a flowchart of a novel method for generating photovoltaic defect data by drone aerial photography according to the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] Please see Figure 1 The steps are as follows:

[0055] Step 1: The drone, equipped with a camera, performs a full-frame scan of the photovoltaic panels along a preset flight path to obtain aerial images (I). raw The attitude sensor acquires the pitch angle θ in real time;

[0056] The flight altitude is dynamically adjusted according to the detection requirements (e.g., 10-100m), and the latitude and longitude coordinates of the photovoltaic panel are recorded in combination with the RTK positioning system to build a spatial information database; the attitude sensor obtains the pitch angle θ in real time for subsequent geometric correction module.

[0057] Step 2: Compare the aerial image I with the flight altitude h and pitch angle θ. raw Correction is performed to obtain the geometrically corrected aerial image I. raw ;

[0058] Step 2 is detailed below:

[0059] Step 21, using the original aerial image I raw The flight altitude h determines the deformation scaling factor s;

[0060] The deformation scaling factor s is calculated as follows:

[0061] s = h / 50, (h∈[10,100]m);

[0062] The deformation scaling factor s increases linearly with the flight altitude h.

[0063] For example, at 50 meters, s = 1, and the deformation is the greatest; when h > 50m, s > 1, the image is magnified to counteract the scale compression effect of high-altitude imaging.

[0064] Step 22, Pitch angle θ, Original aerial image I rawAnd the deformation scaling factor s is embedded in the affine transformation matrix M, for the original aerial image I raw By applying affine transformation, we obtain the geometrically corrected image I. warped The specific calculations are as follows:

[0065] I warped =warpAffine(I raw ,M);

[0066]

[0067] Step 3: Use the current solar panel defect data or take close-up images of defects such as cracks and shading at a distance of 1m using an industrial camera. near Close-up image I near Image I after geometric correction warped The fusion yields the fused chromaticity value C. out With brightness channel L aerial ;

[0068] The purpose is to eliminate color gamut differences between devices.

[0069] The specific steps for step 3 are as follows:

[0070] Step 31, Near-field image I near Aerial images after geometric correction I warped Convert from RGB color space to LAB color space, and separate the luminance (L) and chroma (A / B) channels to obtain the luminance channel L. aerial C, color value of close-up image near and the background chromaticity value C of drone aerial photography aerial ;

[0071] Step 32: Based on the hybrid weighting, adjust the chromaticity values ​​C of the close-up image and the aerial background. aerial Weighted fusion is performed to obtain the fused chromaticity value C. out The specific calculations are as follows;

[0072] C out =α·C near +(1-α)·C aerial ;

[0073] α = 0.9 - 0.2 (h / 10⁻¹);

[0074] α is the blending weight of the LAB channel.

[0075] For example, at a height of 50 meters, α = 0.1, retaining 10% of the chromaticity features of the close-up image and 90% of the aerial background texture. At a height of 10 meters, α = 0.9, retaining 90% of the chromaticity features of the close-up image and 10% of the aerial background texture.

[0076] Step 4: Combine the blended chromaticity values ​​C out With brightness channel L aerial The fused illumination parameters are added to the aerial image brightness channel to obtain a fused image with degraded illumination uniformity. degraded Specifically, it adopts three types of lighting modes.

[0077] Step 41: Define the parameters for the three types of lighting modes.

[0078] Step 42: Set the brightness channel L aerial By integrating the lighting characteristics, the adjusted aerial brightness L′ is obtained, and the specific calculation is as follows:

[0079] L′=β·L aerial +N(0,σ 2 );

[0080] N(0,σ 2 ) represents the Gaussian noise term, β is the brightness gain adjustment, and σ is the degree of dispersion of the light intensity distribution;

[0081] Step 43: Merge LAB images (L′, C out The degraded image I was obtained. degraded The specific calculations are as follows:

[0082] I degraded =Blur(LAB2RGB(L′,C out ),k(h));

[0083] Where k(h)) represents the fuzzy kernel, which is set to 0.5h.

[0084] Noon mode: HSV brightness gain coefficient β = 1.2, variance σ = 0.05;

[0085] Cloudy day pattern: β = 0.4, variance σ = 0.15;

[0086] Dusk mode: β = 0.8, variance σ = 0.1.

[0087] Step 5: Process the degraded image I degraded Perform geometric enhancement and spectral enhancement;

[0088] Geometric enhancements include random rotation, horizontal / vertical rotation, and random cropping;

[0089] Spectral enhancement includes random adjustments to saturation and hue.

[0090] Step 6: Use the pre-trained Mask R-CNN model to initially label defective regions in the geometrically enhanced and spectrally enhanced images, and then manually correct any missing or incorrectly labeled regions.

[0091] Saved in COCO format, containing bounding box, overlay, category labels, and metadata such as flight altitude and lighting mode.

[0092] By establishing a highly adaptive LAB color space blending model, the color chromaticity of near-field defects and the texture of aerial background are dynamically fused, eliminating artifacts caused by color gamut differences between devices, and solving the color gamut offset problem between aerial images and near-field defect data. The generated image is superior to the traditional RGB fusion method.

[0093] A dynamic geometric deformation algorithm based on flight altitude and pitch angle is proposed. Image distortion is corrected by affine transformation matrix, and a height-related scaling factor (s = h / 50) is introduced to compensate for scale compression, thereby compensating for the geometric distortion and scale compression effect of high-altitude imaging.

[0094] A multi-time-period illumination model (noon, cloudy, dusk) is constructed. By combining HSV brightness channel adjustment and Poisson fusion algorithm, the illumination intensity distribution and scattering effect of real scenes are simulated to improve the physical authenticity and environmental adaptability of defect data.

[0095] 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 will 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. A novel method for generating photovoltaic defect data from drone aerial photography, characterized in that: Includes the following steps: Step 1: The drone, equipped with a camera, scans the entire photovoltaic panel along a preset flight path to obtain aerial images. The attitude sensor acquires the pitch angle in real time. , Step 2, Set flight altitude h and pitch angle Aerial images Correction is performed to obtain geometrically corrected aerial images. ; Step 3: Use the current solar panel defect data or take close-up images of defects such as cracks and shading using an industrial camera at a distance of 1 meter. Close-up image Image after geometric correction The fusion yields the fused chromaticity values. With brightness channel ; Step 4: Combine the blended chromaticity values With brightness channel The image with fused illumination parameters into the aerial image brightness channel is a fused image with degraded illumination uniformity. Specifically, it adopts three types of lighting modes; Step 5: Process the degraded image Perform geometric enhancement and spectral enhancement; Step 6: Use the pre-trained Mask R-CNN model to initially label defective regions in the geometrically enhanced and spectrally enhanced images, and then manually correct any missing or incorrectly labeled regions.

2. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 1, characterized in that, Step 2 is detailed below: Step 21: Using the original aerial images Flight altitude h Determine the deformation scaling factor s ; Deformation scaling factor s The specific calculations are as follows: ； Step 22, Pitch Angle Original aerial images and deformation scaling factor s Embedded affine transformation matrix For the original aerial images By applying affine transformation, the geometrically corrected image is obtained. The specific calculations are as follows: ； 。 3. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 2, characterized in that, The specific steps for step 3 are as follows: Step 31, Near-field image and geometrically corrected aerial images Convert from RGB color space to LAB color space, and separate the luminance (L) and chroma (A / B) channels to obtain the luminance channel. Chromaticity values ​​of close-up images and drone aerial photography background chromaticity value ; Step 32: Based on the hybrid weighting, adjust the chromaticity values ​​of the close-up image and the aerial background. Perform weighted fusion to obtain the fused chromaticity values. .

4. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 3, characterized in that, The specific calculations are as follows: ； ； The mixed weights for the LAB channels.

5. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 4, characterized in that, Step 4 specifically adopts three types of lighting modes.

6. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 5, characterized in that, The specific steps for step 4 are as follows: Step 41: Define the parameters for the three types of lighting modes; Step 42: Set the brightness channel By combining lighting characteristics, the adjusted aerial photography brightness is obtained. The specific calculations are as follows: ； Represents the Gaussian noise term. To adjust the brightness gain, Controlling the dispersion of light intensity distribution; Step 43: Merge LAB images ( , ) Obtain degraded image The specific calculations are as follows: ； in, This represents the fuzzy kernel, set to 0.5h.

7. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 6, characterized in that, Noon mode: The specific parameters for the three lighting modes are as follows: HSV Brightness Gain Coefficient =1.2, variance =0.05; Cloudy mode: =0.4, variance =0.15; Dusk Mode: =0.8, variance =0.

1.

8. The novel method for generating photovoltaic defect data by drone aerial photography according to claim 7, characterized in that, Geometric enhancements include random rotation, horizontal / vertical rotation, and random cropping; Spectral enhancement includes random adjustments to saturation and hue.

Images