Two-stage photovoltaic support recognition statistical method and system based on improved SSD model

Abstract

The invention belongs to the technical field of photovoltaic inspection, and particularly provides a two-stage photovoltaic support recognition statistical method and system based on an improved SSD model, and the method comprises the steps: replacing a VGG skeleton in the SSD model with ResNet-50 to form the improved SSD model, carrying out the recognition of a to-be-processed unmanned aerial vehicle orthoimage based on the improved SSD model, obtaining the bounding box coordinates of all photovoltaic strings, and carrying out the recognition of the to-be-processed unmanned aerial vehicle orthoimage. Removing overlapped bounding boxes based on a non-maximum suppression algorithm, and generating an ROI region set; the method comprises the steps of preprocessing an ROI image, extracting all closed contours in the ROI image based on Canny operator edge detection, extracting the closed contours in an edge image and calculating the size of a contour circumscribed rectangle, setting a screening threshold, counting the number of contours conforming to the threshold, accumulating the number of supports of all ROI areas, and obtaining the total number of photovoltaic supports of the whole image. According to the method and system, the number of the photovoltaic supports is identified and counted, and the problem that photovoltaic station construction and operation and maintenance management depend on manual investigation is solved.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic inspection technology, specifically, it relates to a two-stage photovoltaic support identification and statistical method and system based on an improved SSD model. Background Technology

[0002] As the core supporting structure of photovoltaic power plants, photovoltaic (PV) brackets are traditionally used in PV plant construction and operation and maintenance management. This involves manually conducting on-site surveys and counting the number of PV strings and panel brackets to estimate construction progress. However, this traditional manual method is inefficient and highly susceptible to environmental influences, making it unsuitable for the digital construction and operation and maintenance management needs of large-scale PV power plants.

[0003] In existing technologies, single-stage deep learning methods (such as the original YOLO and Faster R-CNN) are often used to directly detect supports. This easily leads to confusion between the hierarchical relationship between "photovoltaic strings" and "individual supports." Furthermore, the small size and dense arrangement of supports result in high detection overlap and a high rate of missed detections. On the other hand, using purely traditional image processing (such as threshold segmentation + contour matching) relies on manually designed features and cannot adapt to changes in lighting (such as oblique sunlight at dawn and dusk) or shadow interference (such as vegetation / equipment occlusion), resulting in low recognition accuracy in complex scenes. Therefore, there is an urgent need for a two-stage method that combines the strong feature extraction capabilities of deep learning with the fine recognition processing of traditional visual algorithms to solve the problem of the lack of mature, high-precision, and fast feasible solutions for photovoltaic power plant construction and operation and maintenance management that rely on manual surveys. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a two-stage photovoltaic support identification and statistical method and system based on an improved SSD model, which identifies and counts the number of photovoltaic supports, and solves the problem of relying on manual surveys for the construction and operation and maintenance management of photovoltaic power plants.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a two-stage photovoltaic support identification and statistical method based on an improved SSD model, comprising the following steps: Phase 1: Photovoltaic string detection and ROI extraction based on the improved SSD model: The VGG skeleton in the SSD model is replaced with ResNet-50 to form an improved SSD model. Based on the improved SSD model, the orthophotos of the UAV to be processed are identified to obtain the bounding box coordinates of all photovoltaic strings. Then, the overlapping bounding boxes are removed based on the non-maximum suppression algorithm to generate a set of ROI regions. Phase Two: Identification and Quantity Count of Photovoltaic Supports within the ROI Area: The ROI image is preprocessed. All closed contours in the ROI image are extracted using the Canny edge detection operator. The closed contours in the edge map are extracted, and the size of the bounding rectangle of each contour is calculated. A filtering threshold is set, and the number of contours meeting the threshold is counted. Add up the number of supports in all ROI regions. This gives the total number of photovoltaic supports in the entire image.

[0006] In a preferred embodiment, the first stage includes the following steps: S101. Improved SSD model architecture design: Replace the VGG skeleton in the SSD model with ResNet-50; S102, Dataset Construction and Model Training: Collect image datasets of photovoltaic strings under different conditions, and then train the improved SSD model constructed in step S101. S103. Photovoltaic string detection and ROI extraction: Input the UAV orthophoto image to be processed into the trained improved SSD model, and output the bounding box coordinates (x1, y1, x2, y2) of all photovoltaic strings. Based on the non-maximum suppression algorithm, overlapping bounding boxes are removed, and a set of ROI regions {R1, R2, ..., R...} is generated. n Each ROI corresponds to an image region of a photovoltaic string in the image.

[0007] In a preferred embodiment, step S102 involves collecting image datasets of photovoltaic strings under different conditions, including photovoltaic string datasets from different seasons, time periods, and weather conditions.

[0008] In a preferred embodiment, in step S102, after collecting image datasets of photovoltaic strings under different conditions, the collected image data is augmented using the Mosaic data augmentation method.

[0009] In a preferred embodiment, the second stage includes the following steps: S201. Preprocess the ROI image; S202. Perform a combination of opening and closing operations on the preprocessed image to remove noise. S203. Extract all closed contours in the ROI image based on Canny operator edge detection; S204. Support Structure Outline Filtering and Quantity Counting: Traverse all edges in the image, extract closed contours from the edge map, calculate the size of the bounding rectangle of the closed contour, set a filtering threshold based on the actual size of the photovoltaic support structure, and count the number of contours that meet the threshold. This refers to the number of photovoltaic (PV) brackets within a single ROI region; the total number of brackets across all ROI regions is calculated by summing them up. This gives the total number of photovoltaic supports in the entire image.

[0010] In a preferred embodiment, step S201, the preprocessing of the ROI image includes the following steps: 1) Size normalization: Each ROI region is scaled to a fixed size, bilinear interpolation is used to maintain the image aspect ratio, and blank areas are filled with gray. 2) HSV Threshold Filtering for Shadow Separation: The normalized RGB image is converted to the HSV color space. Based on the color characteristics of the photovoltaic support, a threshold range is set to filter shadow areas and retain high-brightness non-shadow parts. The formula is as follows: ; in, For pixels, These are the filtered pixel values, where 255 represents the retained candidate bracket area, and 0 represents the background or shadow area. This is the hue value of that pixel. , These are the upper and lower limits of the hue threshold, respectively. This represents the saturation value of that pixel. , These are the upper and lower limits of saturation values, respectively. This represents the brightness value of that pixel. , These are the upper and lower limits of the brightness value, respectively. 3) Image Binarization: Set a grayscale threshold T. If the grayscale value of a pixel in the image is greater than or equal to T, the pixel value is set to 255. In this case, the pixel in the image is classified as the foreground, which includes the photovoltaic strings and the edges of the support structure. Otherwise, the pixel value is set to 0, which includes the background, which includes the part of the ROI region excluding the photovoltaic strings. The expression is: .

[0011] In a preferred embodiment, step S202 involves performing a combination of opening and closing operations on the preprocessed image to remove noise, as follows: 1) Opening operation: A 3×3 rectangular structural element is used for erosion followed by dilation to remove dust and small-area noise from the support surface. The formula is as follows: ; 2) Closing operation: A 5×5 rectangular structural element is first expanded and then eroded to fill the small holes within the bracket outline. The formula is as follows: ; in: The corrosion formula is expressed as: ; The expansion formula is expressed as: .

[0012] In the above formula, A is the processing object in the original image, and B is the rectangular structural element of the operation rule. The opening operation is a 3×3 rectangular structural element, and the closing operation is a 5×5 rectangular structural element.

[0013] In a preferred embodiment, step S203, which involves extracting all closed contours in the ROI image based on Canny operator edge detection, includes the following steps: 1) Gaussian filtering: By convolving the image with a two-dimensional Gaussian kernel, high-frequency noise is removed, reducing the interference of noise on edge detection. The two-dimensional Gaussian kernel function is: ; Where x and y are the image pixel coordinates, Let be the standard deviation of the Gaussian distribution. To control the smoothness of the Gaussian filter, take . ; 2) Gradient calculation: The Gaussian filtered image is compared with the horizontal gradient kernel. Vertical gradient kernel Convolution yields the horizontal gradient. and vertical gradient By calculating the gradients of the image in the horizontal (x) and vertical (y) directions, the magnitude and direction of the edges can be obtained, including the horizontal gradient. and vertical gradient The calculation formula is: ; in, For horizontal gradient kernel, For vertical gradient kernel, This is the image after Gaussian filtering; gradient magnitude The calculation formula is: ; The formula for calculating the gradient direction is: ; in, The gradient direction; (3) Non-maximum suppression: The gradient direction is quantized into 4 main directions, local gradient maximum is retained, and the support edge is thinned to a single pixel width; (4) Lag threshold: Set high and low thresholds to distinguish between strong edges and weak edges, connect weak edges to strong edges to form complete edges, and finally output the binary image of the support edge.

[0014] In the preferred embodiment, in step S204, the dimensions of the circumscribed rectangle of the closed contour are: length ,Width ,area and aspect ratio .

[0015] The present invention also provides a two-stage photovoltaic bracket identification and statistical system based on an improved SSD model, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the two-stage photovoltaic bracket identification and statistical method based on an improved SSD model described above.

[0016] The present invention provides a two-stage photovoltaic support identification and statistical method and system based on an improved SSD model, which has the following beneficial effects: 1. Through a two-stage hierarchical processing approach, the entire process from coarse positioning to fine identification of photovoltaic (PV) brackets is automated, accurately distinguishing the hierarchical relationship between PV strings and individual brackets. The first stage focuses on the overall detection of PV strings using an improved SSD model, outputting bounding boxes and extracting the Region of Interest (ROI), avoiding the confusion between strings and brackets caused by traditional single-stage models directly detecting small-sized brackets. The second stage performs fine bracket identification only within the ROI, significantly reducing the processing range and minimizing background interference (such as vegetation and equipment around the site). This hierarchical design logically eliminates interference from non-string areas on bracket identification, solving the core pain points of high detection overlap and high false negative rates caused by ambiguous hierarchical relationships in existing technologies, making the identification process more closely aligned with the actual physical structure of PV string-bracket configurations.

[0017] 2. Replace the original VGG skeleton of SSD with ResNet-50. Use the residual connection structure of ResNet-50 to solve the gradient vanishing problem in deep networks, and enhance the feature extraction capability for small-sized, densely arranged photovoltaic strings. It is especially suitable for string targets photographed at a distance in UAV orthophotos. 3. The second phase addresses common issues in photovoltaic power plants such as uneven lighting, shading, and surface noise (dust, stains), and designs a multi-dimensional pretreatment process: Converting RGB images to HSV color space, and by setting the H (hue), S (saturation), and V (brightness) threshold ranges for photovoltaic brackets (mostly white / light-colored), the shadow areas projected by vegetation and equipment are precisely filtered out, while retaining high-brightness bracket candidate areas, thus solving the problem of shadow misidentification caused by oblique sunlight in the early morning and late evening. The ROI area is scaled to a fixed size, and bilinear interpolation is used to maintain the proportion, avoiding fluctuations in the bracket size caused by differences in the shooting distance of the photovoltaic strings; grayscale threshold T binarization is used to completely separate the photovoltaic strings from the background and enhance the edge features of the bracket. By combining opening and closing operations, pixel-level dust and noise on the support surface are first removed by opening operations, and then small holes at the support connection are filled by closing operations, ensuring the integrity and continuity of the support outline and providing a clear feature basis for subsequent edge detection.

[0018] 4. By optimizing the Canny operator process, the problems of broken edges and many false edges in traditional edge detection are solved, the interference of noise on edge detection is reduced, and high and low thresholds are set. Strong edges are directly retained, while weak edges are only retained when they are connected to strong edges. This ensures that broken support edges (such as slightly occluded areas) can be completely connected, avoiding missed detection of supports due to edge breaks.

[0019] 5. During the bracket count stage, based on the actual physical dimensions of the photovoltaic brackets, non-bracket interference is effectively eliminated by filtering the geometric features of the circumscribed rectangle of the closed contour. The number of contours that meet the threshold in all ROI areas is accumulated, realizing accurate calculation from single string count to total accumulation across the entire site, avoiding the counting errors of traditional manual statistics.

[0020] 6. This invention is suitable for large-scale and complex photovoltaic power plants, replacing manual labor to improve operation and maintenance efficiency and engineering value. It provides high-precision data support for digital construction (verification of the number of brackets installed) and operation and maintenance (statistics on bracket damage) of photovoltaic power plants, replacing manual surveys, reducing the risks of personnel working in the field, and has significant economic and safety value. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0023] like Figure 1 As shown, a two-stage statistical method for identifying photovoltaic support structures based on an improved SSD model includes the following steps: Phase 1: Photovoltaic string detection and ROI extraction based on the improved SSD model: The VGG skeleton in the SSD model is replaced with ResNet-50 to form an improved SSD model. Based on the improved SSD model, the orthophotos of the UAV to be processed are identified to obtain the bounding box coordinates of all photovoltaic strings. Then, the overlapping bounding boxes are removed based on the non-maximum suppression algorithm to generate a set of ROI regions.

[0024] The first stage includes the following steps: S101. Improved SSD Model Architecture Design: SSD is a convolutional neural network model commonly used for object detection. It performs dense sampling at different locations in an image, extracts features from each level using convolutional layers, and then directly performs classification and regression. To make the SSD model more adaptable to detecting photovoltaic strings, this patent improves the backbone network structure of the SSD model, replacing the original VGG backbone with ResNet-50, thus improving the robustness of the SSD model to small targets.

[0025] S102. Dataset Construction and Model Training: Collect image datasets of photovoltaic strings under different conditions, and then train the improved SSD model constructed in step S101.

[0026] Specifically, a dataset of photovoltaic string data covering different seasons, time periods (differences in shadow length between morning and evening), and weather conditions (cloudy / light rain / dust coverage) was collected. Mosaic data augmentation (randomly stitching together 4 images) was used to improve small target detection performance. During training, a cosine annealing strategy was used to dynamically adjust the learning rate.

[0027] S103. Photovoltaic String Detection and ROI Extraction: Image data of large-scale photovoltaic power plants is collected using drones. The drone orthophotos to be processed are input into a trained improved SSD model, which outputs the bounding box coordinates (x1, y1, x2, y2) of all photovoltaic strings. Overlapping bounding boxes are removed based on the non-maximum suppression algorithm, generating a set of ROI regions {R1, R2, ..., R...}. n Each ROI corresponds to an image region of a photovoltaic string in the image.

[0028] Phase Two: Identification and Quantity Count of Photovoltaic Supports within the ROI Area: The ROI image is preprocessed. All closed contours in the ROI image are extracted using the Canny edge detection operator. The closed contours in the edge map are extracted, and the size of the bounding rectangle of each contour is calculated. A filtering threshold is set, and the number of contours meeting the threshold is counted. Add up the number of supports in all ROI regions. This gives the total number of photovoltaic supports in the entire image.

[0029] The second stage specifically includes the following steps: S201. Preprocess the ROI image.

[0030] Image preprocessing includes operations such as image binarization, size normalization, and HSV thresholding to enhance features and improve accuracy.

[0031] Includes the following steps: 1) Size normalization: Each ROI region is scaled to a fixed size, bilinear interpolation is used to maintain the image ratio, and blank areas are filled with gray to avoid threshold failure caused by size differences in subsequent processing.

[0032] 2) HSV Threshold Filtering for Shadow Separation: Convert the normalized RGB image to the HSV color space. The specific steps are as follows: ① RGB value normalization. The red (R), green (G), and blue (B) values ​​of each pixel in an RGB image are normalized to the 0-1 range using the following formula: ; And calculate the maximum value, minimum value, and their difference: ; ; - .

[0033] ② Calculate the V (lightness) value. Lightness is the maximum value among the RGB components: .

[0034] ③ Calculate the S (saturation) value. Saturation depends on the dispersion of the RGB components, as expressed by the following formula: .

[0035] ④ Calculate the H (hue) value. Hue is expressed as an angle value (0°~360°).

[0036] The calculation of hue depends on which component is... To determine the position of the dominant hue on the color wheel: ; like =0, the image has no hue and is grayscale, at this time H=0.

[0037] After converting the RGB image to the HSV color space using the above calculation method, shadows can be separated using the following method: Based on the color characteristics of the photovoltaic support (mostly white), a threshold range is set to filter out shadow areas and retain high-brightness non-shadow parts. The formula is as follows: ; in, For pixels, These are the filtered pixel values, where 255 represents the retained candidate bracket area, and 0 represents the background or shadow area. This is the hue value of that pixel. , These are the upper and lower limits of the hue threshold, respectively. This represents the saturation value of that pixel. , These are the upper and lower limits of saturation values, respectively. This represents the brightness value of that pixel. , These are the upper and lower limits of the brightness value, respectively.

[0038] 3) Image Binarization: Set a grayscale threshold T. If the grayscale value of a pixel in the image is greater than or equal to T, the pixel value is set to 255. In this case, the pixel in the image is classified as the foreground, which includes the photovoltaic strings and the edges of the support structure. Otherwise, the pixel value is set to 0, which includes the background, which includes the part of the ROI region excluding the photovoltaic strings. The expression is: .

[0039] S202. Perform a combination of opening and closing operations on the preprocessed image to remove noise.

[0040] The operation is as follows: 1) Opening operation: A 3×3 rectangular structural element is used for erosion followed by dilation to remove dust and small-area noise (such as pixel-level impurities) from the support surface. The formula is expressed as: ; 2) Closing Operation: A 5×5 rectangular structural element is first expanded and then eroded to fill small holes (such as gaps at the joints) within the bracket outline, ensuring the integrity of the outline. The formula is expressed as: ; in: The corrosion formula is expressed as: ; The expansion formula is expressed as: .

[0041] In the above formula, A is the processing object in the original image, and B is the rectangular structural element of the operation rule. The opening operation is a 3×3 rectangular structural element, and the closing operation is a 5×5 rectangular structural element.

[0042] S203. Extract all closed contours in the ROI image based on edge detection using the Canny operator.

[0043] Edge detection based on the Canny operator extracts all closed contours in a ROI image through steps such as Gaussian filtering, gradient calculation, non-maximum suppression, and hysteresis threshold determination.

[0044] Includes the following steps: 1) Gaussian filtering: By convolving the image with a two-dimensional Gaussian kernel, high-frequency noise is removed, reducing the interference of noise on edge detection; its two-dimensional Gaussian kernel function is: ; Where x and y are the image pixel coordinates, Let be the standard deviation of the Gaussian distribution. To control the smoothness of the Gaussian filter, it is usually taken as . .

[0045] 2) Gradient calculation: The Gaussian filtered image is compared with the horizontal gradient kernel. Vertical gradient kernel Convolution yields the horizontal gradient. and vertical gradient By calculating the gradients of the image in the horizontal x and vertical y directions, the intensity (magnitude) and direction (angle) of the edge are obtained, including the horizontal gradient. and vertical gradient The calculation formula is: ; in, For horizontal gradient kernel, For vertical gradient kernel, This is the image after Gaussian filtering; gradient magnitude The calculation formula is: ; The formula for calculating the gradient angle is: ; in, The gradient direction.

[0046] (3) Non-maximum suppression: The gradient direction is quantized into four main directions (0°, 45°, 90°, 135°), and only the local gradient maximum value is retained to suppress non-edge pixels and form a single-pixel wide edge.

[0047] (4) Lag threshold: Set high and low thresholds to distinguish between strong edges and weak edges, connect weak edges to strong edges to form complete edges, and finally output the binary image of the support edge.

[0048] S204. Stent outline screening and quantity statistics, specifically including the following steps: 1) Contour Extraction: Traverse all edges in the image, extract closed contours from the edge map, and calculate the size of the bounding rectangle of the closed contour. The size of the bounding rectangle of the closed contour is: length ,Width ,area and aspect ratio .

[0049] 2) Geometric threshold screening: Based on the actual dimensions of the photovoltaic support, a screening threshold is set, and the number of contours that meet the threshold is counted. This refers to the number of photovoltaic (PV) brackets within a single ROI region; the total number of brackets across all ROI regions is calculated by summing them up. This gives the total number of photovoltaic supports in the entire image.

[0050] The present invention also provides a two-stage photovoltaic bracket identification and statistics method based on an improved SSD model, comprising: a memory and a processor, wherein the memory and the processor are interconnected, the memory stores computer instructions, and the processor executes the computer instructions to perform the aforementioned two-stage photovoltaic bracket identification and statistics method based on an improved SSD model.

[0051] After applying and testing the method of this invention in the project, the photovoltaic bracket recognition rate is shown in the table below:

[0052] This method automates the entire process from coarse positioning to fine identification of photovoltaic brackets through a two-stage hierarchical processing, solving the core pain points of existing technologies. It can be adapted to the construction and operation and maintenance phases in complex scenarios such as large-scale photovoltaic power plants in the Northwest desert, providing high-precision, high-efficiency, and high-reliability technical support for the intelligent construction and operation and maintenance of photovoltaic power plants, and has significant engineering value and economic benefits.

[0053] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A two-stage statistical method for identifying photovoltaic support structures based on an improved SSD model, characterized in that, Includes the following steps: Phase 1: Photovoltaic string detection and ROI extraction based on the improved SSD model: The VGG skeleton in the SSD model is replaced with ResNet-50 to form an improved SSD model. Based on the improved SSD model, the orthophotos of the UAV to be processed are identified to obtain the bounding box coordinates of all photovoltaic strings. Then, the overlapping bounding boxes are removed based on the non-maximum suppression algorithm to generate a set of ROI regions. Phase Two: Identification and Quantity Count of Photovoltaic Supports within the ROI Area: The ROI image is preprocessed. All closed contours in the ROI image are extracted using the Canny edge detection operator. The closed contours in the edge map are extracted, and the size of the bounding rectangle of each contour is calculated. A filtering threshold is set, and the number of contours meeting the threshold is counted. Add up the number of supports in all ROI regions. This gives the total number of photovoltaic supports in the entire image.

2. The two-stage photovoltaic support identification and statistical method based on the improved SSD model according to claim 1, characterized in that, The first stage includes the following steps: S101. Improved SSD model architecture design: Replace the VGG skeleton in the SSD model with ResNet-50; S102, Dataset Construction and Model Training: Collect image datasets of photovoltaic strings under different conditions, and then train the improved SSD model constructed in step S101. S103. Photovoltaic string detection and ROI extraction: Input the UAV orthophoto image to be processed into the trained improved SSD model, and output the bounding box coordinates (x1, y1, x2, y2) of all photovoltaic strings. Based on the non-maximum suppression algorithm, overlapping bounding boxes are removed, and a set of ROI regions {R1, R2, ..., R...} is generated. n Each ROI corresponds to an image region of a photovoltaic string in the image.

3. The two-stage photovoltaic support identification and statistical method based on the improved SSD model according to claim 2, characterized in that, In step S102, the image datasets of photovoltaic strings under different conditions include photovoltaic string datasets for different seasons, time periods, and weather conditions.

4. The two-stage photovoltaic support identification and statistical method based on the improved SSD model according to claim 2, characterized in that, In step S102, after collecting image datasets of photovoltaic strings under different conditions, the Mosaic data augmentation method is used to augment the collected image data.

5. The two-stage photovoltaic support identification and statistical method based on an improved SSD model according to claim 1, characterized in that, The second stage includes the following steps: S201. Preprocess the ROI image; S202. Perform a combination of opening and closing operations on the preprocessed image to remove noise. S203. Extract all closed contours in the ROI image based on Canny operator edge detection; S204. Support Structure Outline Filtering and Quantity Counting: Traverse all edges in the image, extract closed contours from the edge map, calculate the size of the bounding rectangle of the closed contour, set a filtering threshold based on the actual size of the photovoltaic support structure, and count the number of contours that meet the threshold. This refers to the number of photovoltaic (PV) brackets within a single ROI region; the total number of brackets across all ROI regions is calculated by summing them up. This gives the total number of photovoltaic supports in the entire image.

6. The two-stage photovoltaic support identification and statistical method based on the improved SSD model according to claim 5, characterized in that, In step S201, the preprocessing of the ROI image includes the following steps: 1) Size normalization: Each ROI region is scaled to a fixed size, bilinear interpolation is used to maintain the image aspect ratio, and blank areas are filled with gray. 2) HSV Threshold Filtering for Shadow Separation: The normalized RGB image is converted to the HSV color space. Based on the color characteristics of the photovoltaic support, a threshold range is set to filter shadow areas and retain high-brightness non-shadow parts. The formula is as follows: ； in, For pixels, These are the filtered pixel values, where 255 represents the retained candidate area for the support structure, and 0 represents the background or shadow area. This is the hue value of that pixel. , These are the upper and lower limits of the hue threshold, respectively. This represents the saturation value of that pixel. , These are the upper and lower limits of saturation values, respectively. This represents the brightness value of that pixel. , These are the upper and lower limits of the brightness value, respectively. 3) Image Binarization: Set a grayscale threshold T. If the grayscale value of a pixel in the image is greater than or equal to T, the pixel value is set to 255. In this case, the pixel in the image is classified as the foreground, which includes the photovoltaic strings and the edges of the support structure. Otherwise, the pixel value is set to 0, which includes the background, which includes the part of the ROI region excluding the photovoltaic strings. The expression is: 。 7. The two-stage photovoltaic support identification and statistical method based on the improved SSD model according to claim 5, characterized in that, In step S202, the noise removal operation performed on the preprocessed image by combining opening and closing operations is as follows: 1) Opening operation: A 3×3 rectangular structural element is used for erosion followed by dilation to remove dust and small-area noise from the support surface. The formula is as follows: ； 2) Closing operation: A 5×5 rectangular structural element is first expanded and then eroded to fill the small holes within the bracket outline. The formula is expressed as: ； in: The corrosion formula is expressed as: ; The expansion formula is expressed as: . In the above formula, A is the processing object in the original image, and B is the rectangular structural element of the operation rule. The opening operation is a 3×3 rectangular structural element, and the closing operation is a 5×5 rectangular structural element.

8. The two-stage photovoltaic support identification and statistical method based on the improved SSD model according to claim 5, characterized in that, In step S203, all closed contours in the ROI image are extracted based on Canny operator edge detection, including the following steps: 1) Gaussian filtering: By convolving the image with a two-dimensional Gaussian kernel, high-frequency noise is removed, reducing the interference of noise on edge detection. The two-dimensional Gaussian kernel function is: ； Where x and y are the image pixel coordinates, Let be the standard deviation of the Gaussian distribution. To control the smoothness of the Gaussian filter, take . ; 2) Gradient calculation: The Gaussian filtered image is compared with the horizontal gradient kernel. Vertical gradient kernel Convolution yields the horizontal gradient. and vertical gradient By calculating the gradients of the image in the horizontal (x) and vertical (y) directions, the magnitude and direction of the edges can be obtained, including the horizontal gradient. and vertical gradient The calculation formula is: ； in, For horizontal gradient kernel, For vertical gradient kernel, This is the image after Gaussian filtering; gradient magnitude The calculation formula is: ； The formula for calculating the gradient direction is: ； in, The gradient direction; (3) Non-maximum suppression: The gradient direction is quantized into 4 main directions, local gradient maximum is retained, and the support edge is thinned to a single pixel width; (4) Lag threshold: Set high and low thresholds to distinguish between strong edges and weak edges, connect weak edges to strong edges to form complete edges, and finally output the binary image of the support edge.

9. A two-stage photovoltaic support identification and statistical method based on an improved SSD model according to claim 5, characterized in that, In step S204, the dimensions of the circumscribed rectangle of the closed contour are: length ,Width ,area and aspect ratio .

10. A two-stage photovoltaic bracket identification and statistical system based on an improved SSD model, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform a two-stage photovoltaic bracket identification and statistical method based on an improved SSD model as described in any one of claims 1 to 9.

Images