Photovoltaic panel ash coating detection method and system based on unmanned aerial vehicle inspection picture

Abstract

The invention, which relates to the technical field of photovoltaic panel cleaning detection, discloses a photovoltaic panel dust covering detection method and system based on an unmanned aerial vehicle inspection picture, and the method comprises the steps: obtaining an image of a to-be-detected region in an RGB format, and carrying out the target detection of a photovoltaic panel through an optic neural network model; segmenting a photovoltaic panel image according to a detection result, performing degradation processing to obtain a low-resolution photovoltaic panel image, inputting the low-resolution photovoltaic panel image into the super-resolution model, and outputting a high-resolution photovoltaic panel image; and constructing a dark channel model and an atmospheric scattering model, carrying out defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image, and carrying out gray covering condition classification by constructing and training an L-CNN lightweight classification model. According to the invention, the image resolution is improved through the super-resolution model; haze noise in the image is effectively removed through the dark channel model and the atmospheric scattering model, and the image definition is improved; by means of a lightweight L-CNN classification model, the dust covering condition of the photovoltaic panel is identified and classified, and personal errors are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic panel cleaning detection, and in particular to a photovoltaic panel dust coating detection method and system based on drone inspection images. Background Art

[0002] In recent years, photovoltaics, as the main form of renewable energy, has developed rapidly with its low pollution, high reliability and wide applicability, and the global installed capacity has reached 660GW. However, photovoltaic panels are exposed to the outdoor environment for a long time and are susceptible to dust pollution. Over time, dust accumulation will significantly reduce the photoelectric conversion efficiency of photovoltaic panels and may even cause hot spots or corrosion. At present, the dust accumulation monitoring system of solar panels mainly determines whether cleaning is needed by measuring the power generation efficiency over a period of time or manual inspection. However, this method is difficult to achieve real-time monitoring and cannot provide timely guidance for the cleaning and maintenance of photovoltaic panels while ensuring accuracy and economic benefits.

[0003] Therefore, a photovoltaic panel dust coverage detection method based on UAV inspection images is proposed, which can realize real-time monitoring and timely detect dust accumulation on photovoltaic panels, so as to carry out targeted cleaning and maintenance, thereby improving the working efficiency and service life of photovoltaic panels. Summary of the invention

[0004] In view of the problems of poor real-time performance, low accuracy and insufficient economic benefits in the existing photovoltaic panel dust accumulation monitoring system, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to achieve real-time detection and evaluation of photovoltaic panel dust coating based on drone inspection pictures.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides a photovoltaic panel dust coverage detection method based on drone inspection images, which includes obtaining an image in RGB format of an area to be inspected; performing photovoltaic panel target detection on the RGB format image through a visual neural network model, and segmenting the photovoltaic panel image according to the detection result, performing degradation processing, and obtaining a low-resolution photovoltaic panel image; inputting the low-resolution photovoltaic panel image into a super-resolution model, and outputting a high-resolution photovoltaic panel image; constructing a dark channel model and an atmospheric scattering model, and performing a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image; constructing and training an L-CNN lightweight classification model, classifying the dust coverage condition of the first photovoltaic panel image, and using the classification result as a basis for evaluating the dust coverage condition of the photovoltaic panel, thereby realizing photovoltaic panel dust coverage detection.

[0008] As a preferred solution of the photovoltaic panel dust coverage detection method for the drone inspection image described in the present invention, the super-resolution model includes a shallow feature extraction layer, a sparse mask layer, a feature fusion layer and an image reconstruction layer; the shallow feature extraction layer includes extracting basic features from the input low-resolution photovoltaic panel image; the sparse mask layer includes sparsely representing the basic features and reducing the data dimension; the feature fusion layer includes fusing the basic features after sparse representation; the image reconstruction layer includes converting the fused basic features into a high-resolution photovoltaic panel image; the basic features include a first high-dimensional vector.

[0009] As a preferred solution of the photovoltaic panel dust detection method of the drone inspection image of the present invention, the high-resolution photovoltaic panel image includes the following steps: extracting the first high-dimensional vector from the low-resolution photovoltaic panel image in the shallow feature extraction layer through the first convolution operation, and the specific formula is as follows:

[0010] F1(I L )=max(0,W1*I L +B1);

[0011] Among them, I L is the input low-resolution photovoltaic panel image, F1(·) is the RELU layer operation, W1 is the convolution kernel parameter of the first convolution operation, and B1 is the bias vector of the first convolution operation.

[0012] The first high-dimensional vector is subjected to a second nonlinear mapping through a second convolution operation to output a second high-dimensional vector. The specific formula is as follows:

[0013] F2(I L )=max(0,W2*F1(IL)+B2);

[0014] Among them, F2(I L ) is the second high-dimensional vector output by the quadratic nonlinear mapping, W2 is the convolution kernel parameter of the second convolution operation, and B2 is the bias vector of the second convolution operation.

[0015] The second high-dimensional vector is sparsely represented in the sparse mask layer through the improved sigmoid activation function. The specific formula is as follows:

[0016]

[0017] Among them, F ac (x) is the activation function, x is the input of the activation function, k h is the upper bound parameter in the training process, k l is the lower bound parameter during training, and k is the factor that controls the linearity of the activation function.

[0018] The second highest dimensional vector after sparse representation is fused with several mask block outputs. The specific formula is as follows:

[0019]

[0020] Among them, F3 is the output of the fusion layer, C1 is the 3*3 convolution calculation, C2 is the 1*1 convolution calculation, and F ac (·) is to improve the calculation of activation function, concat(·) is the concatenation operation, is the output of the nth mask block s.

[0021] The fused second highest dimensional vector is input into the image reconstruction layer and converted into the brightness component of the high-resolution photovoltaic panel image through upsampling operation. The specific formula is as follows:

[0022] F o (Y) = W3*F3+B3;

[0023] Among them, F3(Y) is the deconvolution operation, which restores the image through the upsampling process.

[0024] The high-resolution photovoltaic panel image is trained on the network through the loss function, and the specific formula is as follows:

[0025]

[0026] Among them, Θ = {W1, W2, W3, B1, B2, B3} is the network parameter set, n is the batch size of network training, X i is the brightness component image before bicubic interpolation.

[0027] The brightness component of the high-resolution photovoltaic panel image and the image C in YCBCR format b Components and C r The components are merged to generate a high-resolution solar panel image in the YCBCR format; at the same time, the high-resolution solar panel image in the YCBCR format is converted into a high-resolution solar panel image in the RGB format and dehazed.

[0028] As a preferred solution of the photovoltaic panel dust detection method of the drone inspection image of the present invention, the defogging process includes the following steps: using minimum value filtering to take the minimum channel value of the image pixel point in the input RGB format, obtaining a grayscale image, and establishing a dark channel model. The specific formula is as follows:

[0029]

[0030] Among them, D dark (x) is the dark channel, D cis each channel in the RGB format image, c is the R channel, G channel and B channel in the RGB format image, x is the pixel point of the image, Ω(x) is the filtering area centered on x, and h is the pixel point belonging to the Ω(x) filtering area.

[0031] The 0.15% pixels with the highest brightness are selected from the dark channel model, and the pixel point with the largest weighted average value of the R channel, G channel and B channel is found in the RGB format image. The value of this point is used as the atmospheric light component L to construct an atmospheric scattering model; an equation is established based on the atmospheric scattering model. The specific formula is as follows:

[0032] I c (x) = D c (x)t(x)+L c (1-t(x));

[0033] Among them, c is the R channel, G channel and B channel in the RGB format image, I c (x) is the input foggy image, D c (x) is the dehazed image to be obtained, L c is the atmospheric light component, and t(x) is the transmittance.

[0034] The equation is filtered twice, and the transmittance estimation formula is derived through the dark channel model. The specific formula is as follows:

[0035]

[0036] Among them, ω is the modified weight factor, To estimate the transmittance.

[0037] The weight factor ω is introduced to correct the defogging effect, and the transmittance threshold t0 is set;

[0038] Using the optimized transmittance, atmospheric light components and high-resolution photovoltaic panel image, the defogging formula is applied to calculate and obtain the first photovoltaic panel image. The specific formula is as follows:

[0039]

[0040] Among them, D(x) is the defogging image, I(x) is the photovoltaic panel image before defogging, To estimate the transmittance, L is the atmospheric light component and t0 is the minimum threshold set for the transmittance.

[0041] As a preferred solution of the photovoltaic panel dust detection method for the drone inspection image described in the present invention, the photovoltaic panel target detection process includes the following steps: collecting photovoltaic panel images according to different scenes and annotating them; annotating the photovoltaic panel area as photovoltaic panel to generate a photovoltaic panel target detection data set; selecting the YOLO v8 model as the target detection framework, and inputting the photovoltaic panel target detection data set into the YOLO v8 model for training to construct a visual neural network model; inputting the image in RGB format into the trained visual neural network model to output the detection result; according to the detection result, cutting the photovoltaic panel image from the RGB format image, performing degradation processing, and using the degradation model to generate a low-resolution photovoltaic panel image.

[0042] As a preferred solution of the photovoltaic panel dust detection method of the drone inspection image of the present invention, the detection result includes the position coordinate information of the photovoltaic panel in the RGB format image; the degradation processing includes the following steps: setting the magnification and converting the RGB format image into the YCBCR format image; extracting the brightness Y component from the YCBCR format image and performing normalization processing; based on the traditional bicubic interpolation downsampling model, introducing the image degradation model, the specific formula is as follows:

[0043]

[0044] Among them, I L is a low-resolution image, I H is a high-resolution image, Degeneration(·) is the degradation function, δ is the degradation parameter, k d is the downsampling parameter, and g is Gaussian white noise.

[0045] As a preferred solution of the photovoltaic panel dust detection method for the drone inspection image described in the present invention, the construction process of the L-CNN lightweight model includes the following steps: collecting the first photovoltaic panel image, forming a data set and dividing it into a training set, a verification set and a test set, and marking the degree of dust coverage of the photovoltaic panel; the degree of dust coverage includes dust-free, light dust and heavy dust; constructing a standard CNN layer, performing a convolution operation on the input first photovoltaic panel image, and performing shallow feature extraction; based on the shallow feature extraction, deep feature extraction is performed according to the residual structure, and a classification layer is constructed; the residual structure includes a convolution module; the convolution module includes a 3x3 convolution kernel and a batch normalization layer and a ReLU activation function; the classification layer includes a pooling layer, a fully connected layer and a Softmax function; the L-CNN lightweight model is trained through the training set, and the model performance is evaluated through the test set; the first photovoltaic panel image is input into the trained L-CNN classification model, the classification result is output, the graded dust coverage condition is evaluated, and it is determined whether to clean up, so as to realize the dust coverage detection of the photovoltaic panel.

[0046] In the second aspect, an embodiment of the present invention provides a photovoltaic panel dust coverage detection system based on drone inspection images, which includes: an acquisition module for acquiring an image in RGB format of an area to be inspected; a segmentation module for performing photovoltaic panel target detection on the RGB format image through a visual neural network model, and segmenting the photovoltaic panel image according to the detection result, performing degradation processing, and obtaining a low-resolution photovoltaic panel image; a super-resolution model module for inputting the low-resolution photovoltaic panel image into a super-resolution model and outputting a high-resolution photovoltaic panel image; a defogging module for constructing a dark channel model and an atmospheric scattering model, and performing a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image; a detection module for constructing and training an L-CNN lightweight classification model, classifying the dust coverage condition of the first photovoltaic panel image, and using the classification result as a basis for evaluating the dust coverage condition of the photovoltaic panel, thereby realizing photovoltaic panel dust coverage detection.

[0047] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the photovoltaic panel dust coverage detection method in the drone inspection images as described in the first aspect of the present invention are implemented.

[0048] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the photovoltaic panel dust coverage detection method based on drone inspection images as described in the first aspect of the present invention are implemented.

[0049] The beneficial effects of the present invention are as follows: the present invention utilizes a camera mounted on a drone to take aerial photos of photovoltaic panels, thereby avoiding the high cost and low efficiency of manual inspections; accurately segments the photovoltaic panel area through a visual neural network model, and improves the image resolution in combination with a super-resolution model to ensure the accuracy of subsequent detection; the dark channel model and the atmospheric scattering model effectively remove haze noise in the image and improve the image clarity; with the help of a lightweight L-CNN classification model, the dust coverage of the photovoltaic panels is identified and classified, reducing human errors, and providing a quantitative assessment of the degree of dust coverage, which is helpful for timely maintenance and cleaning, and ensuring the efficient operation of the photovoltaic power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0051] Figure 1 This is a flow chart of the photovoltaic panel dust coverage detection method based on drone inspection images in Example 1.

[0052] Figure 2 This is a schematic diagram of the network structure of the L-CNN lightweight classification model of the photovoltaic panel dust coverage detection method based on drone inspection images in Example 1.

[0053] Figure 3 This is a schematic diagram of the results of the photovoltaic panel dust coverage detection method based on the drone inspection image in Example 2. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0056] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0057] Example 1

[0058] Reference Figure 1-2 , which is the first embodiment of the present invention, provides a photovoltaic panel dust coverage detection method based on drone inspection pictures, including:

[0059] S1: Obtain an image in RGB format of the area to be detected.

[0060] Specifically, a drone equipped with a camera is used to take aerial photos of the ground area where the photovoltaic panels are located; the drone flies along a predetermined path to ensure that all photovoltaic panel areas to be inspected are covered, wherein the camera captures images in RGB format and stores them in the drone's built-in storage device for subsequent processing.

[0061] It should be noted that the drone is a small civilian drone, the Raspberry Pi is an ARM-based microcomputer motherboard and an SD card as a memory hard disk, and the captured image is an RGB format image of the ground top view of the photovoltaic panels installed in the area to be inspected.

[0062] S2: Perform photovoltaic panel target detection on the image in RGB format through a visual neural network model, segment the photovoltaic panel image according to the detection result, perform degradation processing, and obtain a low-resolution photovoltaic panel image.

[0063] Furthermore, the process of photovoltaic panel target detection includes the following steps: collecting photovoltaic panel images according to different scenes and annotating them; annotating the photovoltaic panel area as photovoltaic panel to generate a photovoltaic panel target detection data set; selecting the YOLO v8 model as the target detection framework, and inputting the photovoltaic panel target detection data set into the YOLO v8 model for training, and building a visual neural network model; inputting the RGB format image into the trained visual neural network model, and outputting the detection result; according to the detection result, cutting out the photovoltaic panel image from the RGB format image, and performing degradation processing, and using the degradation model to generate a low-resolution photovoltaic panel image.

[0064] It should be noted that the detection results include the position coordinate information of the photovoltaic panel in the RGB format image; 2000 photovoltaic RGB format images were collected, and the collected photovoltaic panel images were annotated using the annotation software Labelimg.

[0065] Specifically, the visual neural network is based on the YOLO v8 model architecture, and its construction includes the following steps: building a Backbone network for image feature extraction, the backbone network includes a convolution module, a C3 module and an SPP module; the C3 module consists of three convolution modules, the stride of the first convolution module is 2, the stride of the second convolution module and the third convolution module is 1, the convolution module in the C3 module adopts a 3x3 convolution module, and a BN layer and a LeakyReLU activation function are added between each convolution module, and the SPP module includes a convolution layer, a pooling layer and a fully connected layer.

[0066] Furthermore, the features extracted by the backbone network are processed through the Neck structure, which is composed of several convolutional layers and fuses features at different levels through upsampling and downsampling operations. A detection layer is constructed to perform target detection on the features processed by the Neck structure, and each detection frame is classified to determine whether it is the target photovoltaic panel. The features are classified using a fully connected layer plus a Softmax function.

[0067] Furthermore, the degradation processing includes the following steps: setting the magnification, wherein the magnification is set to 3 in the present invention, and converting the image in RGB format into an image in YCBCR format; extracting the brightness Y component from the image in YCBCR format and performing normalization processing; introducing an image degradation model based on a traditional bicubic interpolation downsampling model, and the specific formula is as follows:

[0068]

[0069] Among them, I L is a low-resolution image, I H is a high-resolution image, Degeneration(·) is the degradation function, δ is the degradation parameter, k d is the downsampling parameter, and g is Gaussian white noise.

[0070] S3: Input the low-resolution photovoltaic panel image into a super-resolution model to output a high-resolution photovoltaic panel image.

[0071] Specifically, the super-resolution model includes a shallow feature extraction layer, a sparse mask layer, a feature fusion layer and an image reconstruction layer; the shallow feature extraction layer includes extracting basic features from an input low-resolution photovoltaic panel image; the sparse mask layer includes sparsely representing the basic features and reducing the data dimension; the feature fusion layer includes fusing the basic features after sparse representation; the image reconstruction layer includes converting the fused basic features into a high-resolution photovoltaic panel image; the basic features include a first high-dimensional vector.

[0072] Furthermore, the high-resolution photovoltaic panel image comprises the following steps: extracting the first high-dimensional vector from the low-resolution photovoltaic panel image in the shallow feature extraction layer through a first convolution operation, and the specific formula is as follows:

[0073] F1(I L )=max(0,W1*I L +B1);

[0074] Among them, I L is the input low-resolution photovoltaic panel image, F1(·) is the RELU layer operation, W1 is the convolution kernel parameter of the first convolution operation, and B1 is the bias vector of the first convolution operation.

[0075] Furthermore, the first high-dimensional vector is subjected to a second nonlinear mapping through a second convolution operation to output a second high-dimensional vector. The specific formula is as follows:

[0076] F2(I L )=max(0,W2*F1(I L )+B2);

[0077] Among them, F2(I L ) is the second high-dimensional vector output by the quadratic nonlinear mapping, W2 is the convolution kernel parameter of the second convolution operation, and B2 is the bias vector of the second convolution operation.

[0078] Specifically, the second high-dimensional vector is sparsely represented in the sparse mask layer through the improved sigmoid activation function. The specific formula is as follows:

[0079]

[0080] Among them, F ac (x) is the activation function, x is the input of the activation function, k h is the upper bound parameter in the training process, k l is the lower bound parameter during training, and k is the factor that controls the linearity of the activation function.

[0081] Furthermore, the second high-dimensional vector after sparse representation is subjected to a plurality of mask block output fusions. The specific formula is as follows:

[0082]

[0083] Among them, F3 is the output of the fusion layer, C1 is the 3*3 convolution calculation, C2 is the 1*1 convolution calculation, and F ac (·) is to improve the calculation of activation function, concat(·) is the concatenation operation, is the output of the nth mask block s.

[0084] Furthermore, the fused second highest dimensional vector is input into the image reconstruction layer and converted into the brightness component of the high-resolution photovoltaic panel image through upsampling operation. The specific formula is as follows:

[0085] F o (Y) = W3*F3+B3;

[0086] Among them, F3(Y) is the deconvolution operation, which restores the image through the upsampling process.

[0087] Specifically, the high-resolution photovoltaic panel image is trained on the network through a loss function, and the specific formula is as follows:

[0088]

[0089] Among them, Θ = {W1, W2, W3, B1, B2, B3} is the network parameter set, n is the batch size of network training, X i is the brightness component image before bicubic interpolation.

[0090] Furthermore, the brightness component of the high-resolution photovoltaic panel image and the image C in YCBCR format are combined. b Components and C r The components are merged to generate a high-resolution solar panel image in the YCBCR format; at the same time, the high-resolution solar panel image in the YCBCR format is converted into a high-resolution solar panel image in the RGB format and dehazed.

[0091] S4: constructing a dark channel model and an atmospheric scattering model, and performing a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image.

[0092] Specifically, the defogging process includes the following steps: using minimum value filtering to obtain the minimum channel value of the pixel points of the input RGB format image, obtaining a grayscale image, and establishing a dark channel model. The specific formula is as follows:

[0093]

[0094] Among them, D dark (x) is the dark channel, D c is each channel in the RGB format image, c is the R channel, G channel and B channel in the RGB format image, x is the pixel point of the image, Ω(x) is the filtering area centered on x, and h is the pixel point belonging to the Ω(x) filtering area.

[0095] Furthermore, 0.15% of pixels with the highest brightness are selected from the dark channel model, and the pixel point with the largest weighted average value of the R channel, G channel and B channel is found in the RGB format image, and the value of this point is used as the atmospheric light component L to construct an atmospheric scattering model; an equation is established based on the atmospheric scattering model, and the specific formula is as follows:

[0096] I c (x) = D c (x)t(x)+L c (1-t(x));

[0097] Among them, c is the R channel, G channel and B channel in the RGB format image, I c (x) is the input foggy image, D c (x) is the dehazed image to be obtained, L c is the atmospheric light component, and t(x) is the transmittance.

[0098] Furthermore, the equation is filtered twice and the transmittance estimation formula is derived through the dark channel model. The specific formula is as follows:

[0099]

[0100] Among them, ω is the modified weight factor, To estimate the transmittance.

[0101] It should be noted that the atmospheric scattering model formula is subjected to two minimum filterings, and the specific formula is as follows;

[0102]

[0103] in, To estimate the transmittance, D c (h) is the estimated dehazed image.

[0104] According to the prior principle of the dark channel model, the following formula is derived:

[0105]

[0106] Substituting the above equation into the transmittance estimate equation,

[0107]

[0108] The weight factor ω is introduced to correct the defogging effect, and the transmittance threshold t0 is set.

[0109] Specifically, the optimized transmittance, atmospheric light component and high-resolution photovoltaic panel image are used to calculate and obtain the first photovoltaic panel image using the defogging formula. The specific formula is as follows:

[0110]

[0111] Among them, D(x) is the defogging image, I(x) is the photovoltaic panel image before defogging, To estimate the transmittance, L is the atmospheric light component and t0 is the minimum threshold set for the transmittance.

[0112] S5: Build and train an L-CNN lightweight classification model, classify the dust coverage condition of the first photovoltaic panel image, and use the classification result as a basis for evaluating the dust coverage condition of the photovoltaic panel to achieve dust coverage detection of the photovoltaic panel.

[0113] Specifically, the construction process of the L-CNN lightweight model includes the following steps: collecting the first photovoltaic panel image to form a data set and dividing it into a training set, a validation set and a test set, and marking the degree of dust coverage of the photovoltaic panel; constructing a standard CNN layer, performing a convolution operation on the input first photovoltaic panel image, and performing shallow feature extraction; based on the shallow feature extraction, deep feature extraction is performed according to the residual structure, and a classification layer is constructed; the L-CNN lightweight model is trained through the training set, and the model performance is evaluated through the test set; the first photovoltaic panel image is input into the trained L-CNN classification model, the classification result is output, the dust coverage condition is evaluated and graded, and it is determined whether to clean it up, so as to realize dust coverage detection of the photovoltaic panel.

[0114] It should be noted that the degree of dust coverage includes no dust, light dust and heavy dust. If the classification model determines that the degree of dust coverage is heavy dust, it is judged that cleaning is required; the residual structure includes a convolution module; the convolution module includes a 3x3 convolution kernel and a batch normalization layer and a ReLU activation function; the classification layer includes a pooling layer, a fully connected layer and a Softmax function.

[0115] Furthermore, the present embodiment also provides a photovoltaic panel dust coverage detection system based on drone inspection images, including: an acquisition module, used to acquire an image in RGB format of the area to be inspected; a segmentation module, used to perform photovoltaic panel target detection on the RGB format image through a visual neural network model, and segment the photovoltaic panel image according to the detection result, perform degradation processing, and obtain a low-resolution photovoltaic panel image; a super-resolution model module, used to input the low-resolution photovoltaic panel image into the super-resolution model, and output a high-resolution photovoltaic panel image; a defogging module, used to construct a dark channel model and an atmospheric scattering model, and perform a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image; a detection module, used to construct and train an L-CNN lightweight classification model, classify the dust coverage condition of the first photovoltaic panel image, and use the classification result as a basis for evaluating the dust coverage condition of the photovoltaic panel, thereby realizing photovoltaic panel dust coverage detection.

[0116] This embodiment also provides a computer device, which is suitable for the photovoltaic panel dust coverage detection method in drone inspection images, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the photovoltaic panel dust coverage detection method in drone inspection images as proposed in the above embodiment.

[0117] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0118] The present embodiment also provides a storage medium having a computer program stored thereon, which implements the following steps when executed by a processor: acquiring an image in RGB format of the area to be detected; performing photovoltaic panel target detection on the image in RGB format through a visual neural network model, and segmenting the photovoltaic panel image according to the detection result, performing degradation processing, and obtaining a low-resolution photovoltaic panel image; inputting the low-resolution photovoltaic panel image into a super-resolution model, and outputting a high-resolution photovoltaic panel image; constructing a dark channel model and an atmospheric scattering model, and performing a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image; constructing and training an L-CNN lightweight classification model, classifying the dust covering condition of the first photovoltaic panel image, and using the classification result as a basis for evaluating the dust covering condition of the photovoltaic panel, thereby realizing dust covering detection of the photovoltaic panel.

[0119] In summary, the present invention uses a drone equipped with a camera to take aerial photos of photovoltaic panels, avoiding the high cost and low efficiency of manual inspections; the visual neural network model is used to accurately segment the photovoltaic panel area, and the image resolution is improved in combination with the super-resolution model to ensure the accuracy of subsequent detection; the dark channel model and the atmospheric scattering model are used to effectively remove the haze noise in the image and improve the image clarity; with the help of a lightweight L-CNN classification model, the dust coverage of the photovoltaic panels is identified and classified, reducing human errors, and providing a quantitative assessment of the degree of dust coverage, which is helpful for timely maintenance and cleaning, and ensuring the efficient operation of the photovoltaic power station.

[0120] Example 2

[0121] Reference Figure 3 , which is the second embodiment of the present invention, provides a photovoltaic panel dust coverage detection method based on drone inspection images. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0122] Specifically, the photovoltaic panels installed on the roof of a building in a certain area (longitude: 118.82, latitude: 32.09) and the photovoltaic panels deployed on the ground in a certain area (longitude: 117.17, latitude: 39.11) were taken as samples. A total of 2000 photovoltaic panel images were collected by drone aerial photography in the morning and afternoon every day for analysis and processing, and the evaluation of the results of the method of the present invention and traditional CNN, ResNet-50 and VGG-16 was obtained, as shown in Table 1.

[0123] Table 1 Comparison between the present invention and the traditional method

[0124] Treatment CNN ResNet-50 VGG-16 The present invention No treatment 57.28％ 53.14％ 75.23％ 79.52％ Object Detection 57.12％ 69.14％ 77.12％ 84.82％ Object Detection + Super Resolution 60.76％ 72.58％ 81.34％ 90.38％ Object detection + super resolution + dehazing 61.23％ 73.65％ 83.19％ 93.11％

[0125] Furthermore, it can be seen from Table 1 that it mainly shows the influence of different pre-processing methods on the detection accuracy of the dust coating condition of photovoltaic panels. The processing method refers to the pre-processing steps taken on the inspection pictures taken by the drone before the dust coating condition detection. In particular, no processing is to directly use the original drone inspection pictures for the classification model to detect the dust coating condition. By horizontally analyzing Table 1, it can be observed that when the same pre-processing method is adopted, there is a difference in detection accuracy between the present invention and other methods. The vertical analysis reveals the possible improvement effect of adding pre-processing steps on the detection accuracy of the same method.

[0126] Furthermore, Figure 3 As shown, the implementation effect of each technical step of the present invention can be intuitively seen. The RGB format image taken by the drone inspection is processed by target detection, and the image is cut using the detected photovoltaic panel position information to obtain an image of a single photovoltaic panel; the cut image is relatively blurred, and it is difficult to accurately judge the dust coverage; by super-resolution processing of the image, the image information and features are effectively restored, the resolution of the photovoltaic panel image is improved, and the image is clearer; the high-resolution photovoltaic panel image is defogged to reduce the impact of weather on the image, thereby obtaining a more accurate and clear photovoltaic panel image; dust coverage detection is achieved through L-CNN image classification.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A photovoltaic panel dust detection method based on drone inspection images, characterized by: include, Get the image of the area to be detected in RGB format; Performing photovoltaic panel target detection on the RGB format image through a visual neural network model, segmenting the photovoltaic panel image according to the detection result, performing degradation processing, and obtaining a low-resolution photovoltaic panel image; Inputting the low-resolution photovoltaic panel image into a super-resolution model to output a high-resolution photovoltaic panel image; Constructing a dark channel model and an atmospheric scattering model, and performing a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image; An L-CNN lightweight classification model is constructed and trained to classify the dust coverage condition of the first photovoltaic panel image. The classification result is used as the basis for evaluating the dust coverage condition of the photovoltaic panel to realize dust coverage detection of the photovoltaic panel.

2. The photovoltaic panel dust detection method of the drone inspection image according to claim 1, characterized in that: The super-resolution model includes a shallow feature extraction layer, a sparse mask layer, a feature fusion layer and an image reconstruction layer; the shallow feature extraction layer includes extracting basic features from an input low-resolution photovoltaic panel image; the sparse mask layer includes sparsely representing the basic features and reducing the data dimension; the feature fusion layer includes fusing the basic features after sparse representation; the image reconstruction layer includes converting the fused basic features into a high-resolution photovoltaic panel image; the basic features include a first high-dimensional vector.

3. The photovoltaic panel dust detection method of the drone inspection image according to claim 2, characterized in that: The high resolution photovoltaic panel image comprises the following steps: The first high-dimensional vector is extracted from the low-resolution photovoltaic panel image in the shallow feature extraction layer through the first convolution operation. The specific formula is as follows: F1(I L )=max(0,W 1* I L +B1); Among them, I L is the input low-resolution photovoltaic panel image, F1(·) is the RELU layer operation, W1 is the convolution kernel parameter of the first convolution operation, and B1 is the bias vector of the first convolution operation; The first high-dimensional vector is subjected to a second nonlinear mapping through a second convolution operation to output a second high-dimensional vector. The specific formula is as follows: F2(I L )=max(0,W 2* F1(I L )+B2); Among them, F2(I L ) is the second high-dimensional vector output by the quadratic nonlinear mapping, W2 is the convolution kernel parameter of the second convolution operation, and B2 is the bias vector of the second convolution operation; The second high-dimensional vector is sparsely represented in the sparse mask layer through the improved sigmoid activation function. The specific formula is as follows: Among them, F ac (x) is the activation function, x is the input of the activation function, k h is the upper bound parameter in the training process, k l is the lower bound parameter in the training process, and k is the factor that controls the linearity of the activation function; The second highest dimensional vector after sparse representation is fused with several mask block outputs. The specific formula is as follows: Among them, F3 is the output of the fusion layer, C1 is the 3*3 convolution calculation, C2 is the 1*1 convolution calculation, and F ac (·) is to improve the calculation of activation function, concat(·) is the concatenation operation, is the output of the nth mask block s; the fused second highest dimensional vector is input to the image reconstruction layer and converted into the brightness component of the high-resolution photovoltaic panel image through upsampling operation. The specific formula is as follows: <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> o <h2 style=";text-align:left;direction:ltr"> (Y)=W<h2 style=";text-align:left;direction:ltr"> 3* <h2 style=";text-align:left;direction:ltr"> F3+B3; Among them, F3(Y) is the deconvolution operation, which restores the image through the upsampling process; The high-resolution photovoltaic panel image is trained on the network through the loss function, and the specific formula is as follows: Among them, Θ = {W1, W2, W3, B1, B2, B3} is the network parameter set, n is the batch size of network training, X i is the brightness component image before bicubic difference; The brightness component of the high-resolution photovoltaic panel image and the image C in YCBCR format b Components and C r Component merging to generate high-resolution solar panel images in YCBCR format; At the same time, the high-resolution solar panel image in YCBCR format is converted into a high-resolution solar panel image in RGB format and dehazed.

4. The photovoltaic panel dust detection method of the drone inspection image according to claim 3, characterized in that: The defogging process comprises the following steps: The minimum value filter is used to take the minimum channel value of the pixel point of the input RGB format image, obtain the grayscale image, and establish the dark channel model. The specific formula is as follows: Among them, D dark (x) is the dark channel, D c is each channel in the RGB format image, c is the R channel, G channel and B channel in the RGB format image, x is the pixel point of the image, Ω(x) is the filter area centered on x, and h is the pixel point belonging to the Ω(x) filter area; Select the 0.15% pixels with the highest brightness from the dark channel model, and find the pixel point with the largest weighted average value of the R channel, G channel and B channel in the RGB format image. Use the value of this point as the atmospheric light component L to build the atmospheric scattering model. The equation is established based on the atmospheric scattering model. The specific formula is as follows: I c (x)=D c (x)t(x)+L c (1-t(x)); Among them, c is the R channel, G channel and B channel in the RGB format image, I c (x) is the input foggy image, D c (x) is the dehazed image to be obtained, L c is the atmospheric light component, t(x) is the transmittance; The equation is filtered twice, and the transmittance estimation formula is derived through the dark channel model. The specific formula is as follows: Among them, ω is the modified weight factor, To estimate the transmittance; The weight factor ω is introduced to correct the defogging effect, and the transmittance threshold t0 is set; Using the optimized transmittance, atmospheric light components and high-resolution photovoltaic panel image, the defogging formula is applied to calculate and obtain the first photovoltaic panel image. The specific formula is as follows: Among them, D(x) is the defogging image, I(x) is the photovoltaic panel image before defogging, To estimate the transmittance, L is the atmospheric light component and t0 is the minimum threshold set for the transmittance.

5. The photovoltaic panel dust detection method of the drone inspection image according to claim 1, characterized in that: The photovoltaic panel target detection process includes the following steps: Collect and annotate photovoltaic panel images according to different scenarios; The photovoltaic panel area is marked as photovoltaic panel, and a photovoltaic panel target detection dataset is generated; The YOLO v8 model is selected as the target detection framework, and the photovoltaic panel target detection dataset is input into the YOLO v8 model for training to build a visual neural network model; Input the RGB format image into the trained visual neural network model and output the detection results; According to the detection result, a photovoltaic panel image is cut out from the image in RGB format, and degradation processing is performed, and a low-resolution photovoltaic panel image is generated using a degradation model.

6. The photovoltaic panel dust detection method of the drone inspection image according to claim 5, characterized in that: The detection result includes the position coordinate information of the photovoltaic panel in the image in RGB format; the degradation processing includes the following steps: Set the magnification and convert the RGB image to the YCBCR image. Extract the brightness Y component from the image in YCBCR format and perform normalization; Based on the traditional bicubic interpolation downsampling model, an image degradation model is introduced. The specific formula is as follows: Among them, I L is a low-resolution image, I H is a high-resolution image, Degeneration(·) is the degradation function, δ is the degradation parameter, k d is the downsampling parameter, and g is Gaussian white noise.

7. The photovoltaic panel dust detection method of the drone inspection image according to claim 1, characterized in that: The construction process of the L-CNN lightweight model includes the following steps: Collecting the first photovoltaic panel images to form a data set and dividing it into a training set, a validation set and a test set, and marking the degree of gray coverage of the photovoltaic panels; The dust coverage degree includes dust-free, light dust and heavy dust; Constructing a standard CNN layer, performing a convolution operation on the first photovoltaic panel image input, and performing shallow feature extraction; Based on the shallow feature extraction, deep feature extraction is performed according to the residual structure, and a classification layer is constructed; The residual structure includes a convolution module; the convolution module includes a 3x3 convolution kernel, a batch normalization layer and a ReLU activation function; the classification layer includes a pooling layer, a fully connected layer and a Softmax function; The L-CNN lightweight model is trained using the training set, and the model performance is evaluated using the test set. The first photovoltaic panel image is input into the trained L-CNN classification model, and the classification result is output to evaluate and classify the dust coverage status, and determine whether to clean it, so as to realize the dust coverage detection of the photovoltaic panel.

8. A photovoltaic panel dust-covered detection system based on drone inspection images, a photovoltaic panel dust-covered detection method based on drone inspection images as claimed in any one of claims 1 to 7, characterized in that: include, An acquisition module is used to acquire an image in RGB format of the area to be detected; A segmentation module is used to detect photovoltaic panel targets on the image in RGB format through a visual neural network model, and to segment the photovoltaic panel image according to the detection result, and to perform degradation processing to obtain a low-resolution photovoltaic panel image; A super-resolution model module, used for inputting the low-resolution photovoltaic panel image into a super-resolution model and outputting a high-resolution photovoltaic panel image; A defogging module, used for constructing a dark channel model and an atmospheric scattering model, and performing a defogging operation on the high-resolution photovoltaic panel image to obtain a first photovoltaic panel image; The detection module is used to build and train the L-CNN lightweight classification model, classify the dust coverage condition of the first photovoltaic panel image, and use the classification result as the basis for evaluating the dust coverage condition of the photovoltaic panel to realize the dust coverage detection of the photovoltaic panel.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the photovoltaic panel dust coverage detection method based on the drone inspection image described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the photovoltaic panel dust coverage detection method based on drone inspection images described in any one of claims 1 to 7 are implemented.

Images