Label identification method for retired photovoltaic panel

Abstract

The invention relates to the technical field of photovoltaic panels, in particular to a decommissioned photovoltaic panel label identification method, which comprises the following steps: acquiring an image of a decommissioned photovoltaic panel, and constructing a data set; obtaining a character image area in the photovoltaic panel label by using the network detection model, and generating a character textbox; correcting the text box image containing the characters; identifying and expanding lines of the keyword textbox by utilizing a dynamic line threshold value; converting an image coordinate of the expanded keyword textbox into a pixel coordinate, and cutting the expanded keyword textbox according to a boundary value of the pixel coordinate to obtain a sub-image of the expanded keyword textbox; and extracting characters in the sub-images of the expanded keyword textbox by using a character recognition model. According to the method and the device, the problem that the character recognition model cannot accurately recognize the keywords due to the fact that the recognition frames of the keyword tags are serial during extension is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic panels, and in particular to a method for identifying labels of retired photovoltaic panels. Background Art

[0002] Photovoltaic panels that have been in service for a long time have gradually entered the retirement stage and need to be recycled and recycled as resources. Retired photovoltaic panels follow the full life cycle responsibility system, and it is necessary to track information on the circulation and use processes of photovoltaic panels, including production, sales, service, retirement, and disassembly.

[0003] The only information about retired photovoltaic panels is the label on the back panel, which contains key information such as manufacturer, maximum power, production serial number, weight and size. It is necessary to identify the back panel and extract the information on the label.

[0004] Currently, there are the following problems in extracting retired photovoltaic panels:

[0005] 1. The photovoltaic panel labels of different manufacturers vary in design, information content and layout format. There is a series when extending the recognition frame that recognizes the keyword label, which makes the character recognition model unable to accurately recognize the keyword;

[0006] 2. The same key information in the labels of photovoltaic panels produced by different manufacturers, or different batches of photovoltaic panels produced by the same manufacturer, may be expressed in different ways; photovoltaic panel labels have significant differences in layout, expression of the same (or different) information keywords, fonts, font sizes, etc., which can easily lead to confusion or missed detection during intelligent recognition;

[0007] 3. The extension of the recognition frame causes the extraction of irrelevant key information, wasting recognition resources. Summary of the invention

[0008] In view of the shortcomings of the existing methods, the present invention solves the problem that when the recognition box of the keyword label is extended, there is a series, which causes the character recognition model to be unable to accurately recognize the keyword.

[0009] The technical solution adopted by the present invention is: a method for identifying tags of retired photovoltaic panels comprises the following steps:

[0010] Step 1: Collect images of retired photovoltaic panels and build a dataset;

[0011] As a preferred implementation of the present invention, LabelImg is used to label the photovoltaic panel image.

[0012] As a preferred implementation of the present invention, data enhancement preprocessing is performed on the photovoltaic panel image.

[0013] Step 2: Use the network detection model to obtain the character image area in the photovoltaic panel label and generate a character text box;

[0014] As a preferred implementation of the present invention, the network detection model includes: Faster R-CNN, SSD, MaskR-CNN, and RetinaNet.

[0015] Step 3: Correct the image of the text box containing the characters; use the dynamic line threshold to identify and expand the line of the keyword text box;

[0016] As a preferred embodiment of the present invention, using a dynamic row threshold to identify and expand the rows of the keyword text box includes:

[0017] First, count the upper and lower boundaries of all character text boxes of a label and calculate the average height height_avg of all character text boxes;

[0018] Secondly, use height_avg and k to set the line spacing threshold y_threshold = height_avg*k, where k is an empirical coefficient;

[0019] Next, traverse the text boxes of all characters, and calculate whether the absolute values ​​of the differences between the upper and lower boundaries of all non-keyword text boxes and a certain keyword text box are less than or equal to y_threshold; if so, the non-keyword text box is a suspected keyword text box;

[0020] Finally, the expanded keyword text box is obtained by using the text box with the minimum value of the left boundary x1 in the suspected keyword text box and the maximum value of the right boundary x2 in the suspected keyword text box.

[0021] As a preferred implementation of the present invention, a field dependency tree is preset for the keyword and its adjacent fields.

[0022] Step 4: convert the image coordinates of the expanded keyword text box into pixel coordinates, and crop the expanded keyword text box according to the boundary values ​​of the pixel coordinates to obtain a sub-image of the expanded keyword text box;

[0023] As a preferred embodiment of the present invention, the formula for the boundary value of pixel coordinates is:

[0024] x_min=(x_center-width / 2)*W

[0025] x_max=(x_center+width / 2)*W

[0026] y_min=(y_center-height / 2)*H

[0027] y_max=(y_center+heignt / 2)*H

[0028] Among them, x_center, y_center are the normalized center coordinates of the keyword text box; width, height are the normalized width and height of the expanded keyword text box; W, H are the width and height of the original image respectively.

[0029] Step 5: extract characters from the sub-image of the expanded keyword text box using a text recognition model;

[0030] As a preferred implementation of the present invention, the text recognition model is an OCR model.

[0031] As a preferred embodiment of the present invention, a system for identifying tags of retired photovoltaic panels includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a method for identifying tags of retired photovoltaic panels.

[0032] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and when the computer program code is executed by a processor, a method for identifying tags of retired photovoltaic panels is implemented.

[0033] Beneficial effects of the present invention:

[0034] 1. The present invention uses deep learning algorithms to perform target detection and OCR to identify photovoltaic panel labels, and divides label detection and text recognition into different stages; first, the target detection algorithm is used to identify the key information part of the label, and the OCR technique is used to perform text recognition on the detected keyword character image; the photovoltaic panel label is accurately located and the text information therein is identified;

[0035] 2. When expanding, dynamic row threshold calculation and up-and-down row inference mechanism are introduced to accurately determine the character row boundaries in the photovoltaic panel label, improve the adaptability to different label designs, and then reduce the error recognition problem caused by serialization;

[0036] 3. By presetting the field dependency tree, the parameter values ​​corresponding to the keywords to be identified are obtained, which reduces the amount of unnecessary identification keywords, reduces the amount of subsequent ORC identification, and saves identification resources;

[0037] 4. Unify the naming standards of the information recognized by OCR and automatically fill it into the database by field; and then propose a custom module to optimize the information extraction and database entry process, automatically identify and record the label information of each retired photovoltaic panel, making information tracing more convenient and reliable. The system can accurately track the processing status and historical records of each photovoltaic panel, improving the transparency and traceability of information. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of a method for identifying tags of retired photovoltaic panels according to the present invention;

[0039] Figure 2 These are label images from different manufacturers or batches;

[0040] Figure 3 It is a schematic diagram of a keyword text box of the present invention;

[0041] Figure 4 is a schematic diagram of keyword and non-keyword text boxes of the present invention;

[0042] Figure 5 is a schematic diagram of an expanded keyword text box of the present invention;

[0043] Figure 6 It is the OCR recognition effect diagram of the present invention. DETAILED DESCRIPTION

[0044] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0045] like Figure 1 As shown, a method for identifying tags of retired photovoltaic panels includes the following steps:

[0046] Step 1: Collect images of retired photovoltaic panels and build a dataset;

[0047] Collect complete images of the back of retired photovoltaic panels, annotate the photovoltaic panel label (label for short), and manually annotate the keyword character images in the photovoltaic panel labels to construct label datasets and keyword datasets;

[0048] Take high-quality images of actual decommissioned PV panel labels using a high-resolution camera or mobile phone to ensure that the characters and symbols on the labels are clearly captured;

[0049] By finding and visiting storage locations of retired photovoltaic panels, visiting waste photovoltaic panel recycling stations, storage warehouses, and installation sites, more diverse label images can be obtained;

[0050] Collect photovoltaic panels of various brands and models, including well-known brands and regional brands, to ensure exposure to different types of labels; ensure shooting under different lighting and background conditions, and shoot under different lighting conditions such as strong light, weak light, backlight, and shadow to enhance the diversity of the dataset. The image samples include various photovoltaic panels and their labels, ensuring the generalization ability of the model.

[0051] Use the LabelImg tool to label the label area on the back of the photovoltaic panel in the label data set, and then use LabelImg to label the key information in the label area;

[0052] Keywords include: company logo, maximum power, production serial number, weight and size, etc.;

[0053] The annotation information includes the bounding box coordinates of the character image of the key information;

[0054] The same keywords should also include different expressions such as Chinese, English, and logos, and different expressions should be given the same meaning when annotating; Figure 2 For the label images of photovoltaic panels from different manufacturers or different batches of the same manufacturer, the same key information has different expressions on different labels, and the same meaning is given to different expressions when annotated.

[0055] For example, in photovoltaic panel labels, due to different design styles of each company, the company logo may be presented in a variety of ways, including: a combination of a pattern and a company name, such as a graphic + brand pinyin / Chinese characters or directly using the company name (in Chinese and English) as the logo; these company logos presented in different ways all express the same meaning and need to be uniformly classified as logos in labeling and identification;

[0056] Maximum power can be expressed in Chinese or English. In Chinese labels, the maximum power information is usually expressed in several different ways, including: maximum power, maximum output power; in English labels, maximum power is usually also expressed in different ways, common English expressions include: Maximum Power, Peak Power, Rated Maximum Power; also includes Pmax or Pm as its abbreviation, these two different keywords need to be uniformly classified as Pmax;

[0057] The Chinese expression of weight is component weight, and the English expression of weight is Weight; both are uniformly classified as Weight;

[0058] The Chinese expressions of size include component size and size; the English expressions include Size, Dimensions (MM), and Dimensions, which are all collectively classified as Size;

[0059] The format of the production serial number may not be uniform. It may be a combination of letters and numbers, or it may be all numbers. If the serial number format is 123456789, select the entire serial number area and mark it as Serial Number. If the serial number is ABC123456, directly select the part with numbers and letters and mark it as Serial Number.

[0060] The photovoltaic panel labels and annotated information documents are used to construct a data set, and the data set is divided into a training set and a test set in a ratio of 8:2 to complete the construction of the data set. After the annotation is completed, the generated XML format data is saved and converted into TXT format data;

[0061] The photovoltaic panel labels are preprocessed, including data enhancement. The data enhancement methods include random cropping, horizontal flipping, adjusting the hue and saturation of the image, randomly filling some areas with 0 pixel values, rotating at any angle, and adding Gaussian noise and salt and pepper noise.

[0062] Step 2: Use the network detection model to obtain the character image area in the photovoltaic panel label and generate a character text box;

[0063] Network detection models include: YOLOv series, YOLOv8, YOLOv9, etc.; Faster R-CNN, SSD, Mask R-CNN, RetinaNet, etc.;

[0064] Take the YOLOv8 model as an example:

[0065] Use the labeled training set to train YOLOv8 and generate the target detection weight file;

[0066] According to the trained YOLOv8 model, the features of the retired photovoltaic panel images are extracted to obtain the label features. According to the extracted features, the corresponding regions of interest are generated; the intersection-over-union ratio is obtained by calculating the overlapping part between the retired photovoltaic panel label information image and the annotated image; the intersection-over-union ratio is substituted into the loss function to calculate the loss value; back propagation is performed according to the loss value to modify the annotated image until the target annotated image is generated; the intersection-over-union ratio between the retired photovoltaic panel label image and the target annotated image is calculated; the target intersection-over-union ratio is substituted into the loss function to obtain the target loss value; check whether the target loss value is less than or equal to the preset loss value; if so, extract the weight file; if not, continue to modify the target defect annotated image until the conditions are met; the activation function uses the SiLU function, and the calculation formula is as follows:

[0067]

[0068] The loss function calculation formula is as follows:

[0069]

[0070] Among them, L box is the bounding box loss, L cls is the classification loss, L obj is the confidence loss, and ||θ|| is the regularization term. The positive sample is selected based on the weighted score of the classification and regression scores. The formula is:

[0071] k=t α ×h β

[0072] Among them, t is the predicted score corresponding to the labeled category, h is the IoU value between the predicted box and the real box, α and β are weight hyperparameters, and the multiplication of the two can measure the degree of alignment;

[0073] The main indicators for measuring detection accuracy include precision P, recall R, and mean average precision mAP.

[0074] Experimental process:

[0075] The resolution of the sample image is set to 640×640 pixels, the training batch is 16, the initial learning rate is 0.01, and all control models are trained for 300 epochs according to these parameters; the obtained training weights are used to detect the photovoltaic panel labels and obtain the region of interest detection results. The results show that the overall accuracy P of the YOLOv8 model for the region of interest is 91.5%, the recall rate R is 88%, and the mean average precision mAP0.5 is 89%.

[0076] Step 3: Correct the image of the text box containing the characters; use the dynamic line threshold to identify and expand the line of the keyword text box;

[0077] The correction includes: firstly, calculating the tilt angle of the keyword character image area and rotating and correcting it to the horizontal; secondly, using dynamic line threshold calculation and up-down line inference mechanism to accurately determine the character line boundary of the line where the keyword character image area after rotation correction is located;

[0078] The keyword character image area is obtained by using the YOLOv8 model to identify the specific location corresponding to the keyword label and obtain the bounding box of the keyword label area, such as Figure 3 As shown in the figure, the red bounding box of the keyword Pmax indicates the position of the photovoltaic panel label in the image. The coordinates of the red bounding box are used to crop the label area containing the keyword image from the original image. The cropped image is the sub-image containing the photovoltaic panel label. The sub-image is used as the input image for keyword extraction of the subsequent OCR model.

[0079] When the detected keyword character image area has an inclination angle, that is, it is not in a horizontal state; the keyword character image is rotated to a horizontal state, and the frame coordinates of the keyword character image are extracted;

[0080] Rotating the photovoltaic panel label image to horizontal includes: first, using edge detection to extract the edge of the character area; second, using Hough line transform to detect the direction of the longest set of straight lines in the character area, and calculating the inclination angle of the character area based on the angle between the longest straight line and the horizontal line; third, using the inclination angle theta to generate a rotation matrix, (cx, cy) is the center point of the character area, using -theta for counterclockwise rotation, and scale = 1.0 to maintain the original image size; finally, applying the rotation matrix to rotate the entire image counterclockwise to make the text horizontal.

[0081] Figure 3 The identified Pmax and its border position are used as output for subsequent border positioning, cropping and identification; when the image to be processed is input into the network model for detection, the identified key information and its border position are used as output.

[0082] Extensions include:

[0083] Traverse all text box detection results, all text boxes include: keyword text boxes and non-keyword text boxes; search for keywords in the label, and obtain the upper and lower boundaries y1, y2 and left and right boundaries x1, x2 of the coordinate values ​​of the keyword text box, such as Figure 4 ; According to the coordinate information of the keyword text box, expand the start and end positions of its line, such as Figure 5 It is a text box expanded according to the keyword Pmax.

[0084] In order to ensure that all characters in the same line with keywords are extracted, dynamic line threshold calculation and up and down line inference mechanism are introduced to determine the character line boundaries in the label, improving the adaptability to different labels;

[0085] First, all characters, including keywords and non-keywords, count the upper and lower boundaries of all character text boxes of a label, and calculate the average height of the text boxes of all characters:

[0086]

[0087] Where N is the number of character text boxes, y i 1 is the upper boundary of the text box of the i-th character, y i 2 is the lower boundary of the text box of the i-th character.

[0088] The standard deviation height_std is calculated for the upper and lower boundaries of all character text boxes to determine the range of character size variation in the label.

[0089] Secondly, use height_avg and k to set the line spacing threshold y_threshold; k is an empirical coefficient, which is a custom parameter. For example, k is (0-1); the formula of y_threshold is:

[0090] y_threshold = height_avg * k

[0091] Secondly, traverse all the character text boxes and calculate the height y of all non-keyword text boxes respectively i The absolute value of the difference between y1 and the y1 of a keyword text box and y i The absolute value of the difference between y2 and a keyword text box. If the absolute value of the difference is less than or equal to y_threshold, it is defined as a suspected keyword text box.

[0092] For example Figure 4 The upper boundary y of the non-keyword Maximum in the first row 1 1- The absolute value of the difference between the upper boundary y1 of the keyword Pmax and y 1 If the absolute value of the difference between 2 and y2 is less than y_threshold, the text box of Maximum is suspected to be a text box in the same field as the keyword.

[0093] Finally, the text box with the minimum x1 value on the left edge of the suspected keyword text box is used as the starting position of the current row, and the text box with the maximum x2 value on the right edge of the suspected keyword text box is used as the ending position of the current row to obtain the expanded keyword text box;

[0094] like Figure 5 As shown, the starting position of the current row of Pmax is the left boundary x1 of Maximum, and the ending position of the current row is the right boundary x2 of 270W (0~+5W).

[0095] The present invention adapts to labels with different character densities by dynamically updating y_threshold; for areas with denser characters, k can be appropriately increased to adjust the line spacing threshold; for areas with sparser characters, k can be reduced to make line inference more flexible; the character range of the entire line can be accurately extracted by dynamically adjusting y_threshold; possible misjudgments can be corrected by adjusting the threshold, for example, to prevent different character text boxes from being mistakenly identified as the same line, or to prevent some characters in the same line from being missed.

[0096] Preferably, the expanded keyword text box may expand two non-related keywords together, increasing the workload of subsequent OCR model recognition, such as Figure 5As shown, Pmax and its corresponding power value W should be in the same expanded keyword text box to be the keywords and parameter values ​​that need to be processed; if Pmax and model are in the same expanded keyword text box, it is an invalid keyword-parameter value pair;

[0097] The present invention uses a preset field dependency tree to structurally represent the preset relationship between a keyword field and its adjacent fields; for example, the word "Pmax" is adjacent to the power value field; "Company Name" is generally located in the top area of ​​the label; after finding the field associated with the keyword, the field content is verified to ensure its logical relationship and integrity; for example, ensure that the power value is immediately on the right side of Pmax and is in a reasonable position; if an extended keyword text box that meets the conditions is found, the content is merged with the keyword to form a complete field, for example, if the right side of Maximum Power (Pmax) is 270W (0~+5W), the logical relationship is correct, if the right side of Maximum Power (Pmax) is the manufacturer, the logical relationship is wrong, and the extended keyword text box should be discarded; this prevents misjudgment or missed judgment and reduces the recognition amount of subsequent OCR models.

[0098] Step 4: convert the image coordinates of the expanded keyword text box into pixel coordinates, and cut the expanded keyword text box according to the boundary value of the pixel coordinates to obtain a sub-image of the expanded keyword text box;

[0099] The expanded keyword text box is cropped out from the original image to generate a series of sub-images containing the expanded keyword text box.

[0100] Converting the image coordinates of the expanded keyword text box to pixel coordinates includes:

[0101] Get the size of the original image as (W, H) (width and height), normalize the coordinates of the expanded keyword text box, and get the center point coordinates (x_center, y_center). Width is the width of the expanded keyword text box, normalized to the ratio relative to the image width; height is the height of the expanded keyword text box, normalized to the ratio relative to the image height. The normalized width and height define the ratio of the size of the bounding box to the image size; the formula for the corresponding pixel coordinate boundary value is:

[0102] x_min=(x_center-width / 2)*W

[0103] x_max=(x_center+width / 2)*W

[0104] y_min=(y_center-height / 2)*H

[0105] y_max=(y_center+heignt / 2)*H

[0106] The normalized coordinates are converted into pixel coordinates to crop out the expanded keyword text box sub-image; cropping is performed according to the pixel coordinate boundary value to obtain a series of sub-images containing the expanded keyword text box; cropping is performed using the pixel coordinate boundary value to ensure that the cropping target is within the image boundary.

[0107] OCR is used for text recognition. The expanded keyword character images obtained by positioning are processed by OCR to extract the character information in the label. OCR recognition uses CRNN network. CRNN contains three layers: convolution layer, recurrent layer and transcription layer. Since the length of English words in each image is inconsistent, but the feature length extracted after CNN is certain, a transcription layer is required to obtain the final result.

[0108] In the prediction process, the features of the text image are first extracted through a standard CNN network, and then the feature vectors are fused using a bidirectional long short-term memory network (BLSTM) to capture the contextual information of the character sequence; then, the model calculates the probability distribution for each column of features, and finally predicts the text sequence through connectionist temporal classification (CTC);

[0109] The CTC loss function is defined as the negative maximum likelihood function of probability, and the formula is:

[0110]

[0111] Where x' is the input feature vector sequence; y' is the target label sequence, which may include blank labels; B(t) is the set of paths that generate the target sequence t from all possible label sequences; P(y'|x') is the probability of generating the label sequence y' given the input sequence x'; CTC loss optimization can help the model learn effective feature representations, thereby improving performance in irregular sequence tasks.

[0112] like Figure 6 , is a schematic diagram of keywords and non-keywords after OCR recognition, and 0.965 and 0.943 are the confidence levels respectively.

[0113] In addition, the keywords output by the OCR model, such as power, weight, and size, may be expressed in multiple ways, such as Chinese and English, and in different units. For example, the power field may be Pmax or maximum power, the weight unit may be g or kg, and the size unit may be mm or cm; for power, weight, size and other fields, use rules and regular expressions to unify different expressions; for maximum power, unify the field as Maximum Power, the value is unified in W unit; for weight, the unified field is Weight, and the value is unified in kg; for size, the unified field is Dimension, and the value is unified in cm; for the production serial number, it is usually a combination of letters and numbers. Check whether it contains only English letters and numbers, and store it directly without unit conversion; format the text information obtained by OCR recognition into a standardized data format, design the database table structure to store the label information of retired photovoltaic panels, and use MySQL as the database. Enter the formatted label information into the database for storage; when searching for information and filling in the database, the information to be identified and the corresponding value will be divided into two lines, and a custom module will be used. The first traversal is used to extract information, and the second traversal checks the keyword and searches for the value after the keyword; if the keyword is found in the current line, check the next line to obtain the corresponding value. By checking the current line and the next line, ensure that the field information divided into multiple lines can be obtained; when filling in the database, check whether the field is a production serial number, define a regular expression to match the production serial number, and for other fields, use string segmentation to extract the corresponding value.

[0114] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for identifying tags of retired photovoltaic panels, characterized in that: The following steps are involved: Step 1: Collect images of retired photovoltaic panels and build a dataset; Step 2: Use the network detection model to obtain the character image area in the photovoltaic panel label and generate a character text box; Step 3: Correct the image of the text box containing the characters; use the dynamic line threshold to identify and expand the line of the keyword text box; Step 4: convert the image coordinates of the expanded keyword text box into pixel coordinates, and crop the expanded keyword text box according to the boundary values ​​of the pixel coordinates to obtain a sub-image of the expanded keyword text box; Step 5: Use the text recognition model to extract characters from the sub-image of the expanded keyword text box.

2. The method for identifying tags of retired photovoltaic panels according to claim 1, characterized in that: Using dynamic row thresholds to identify and expand rows in keyword text boxes includes: First, count the upper and lower boundaries of all character text boxes of a label and calculate the average height height_avg of all character text boxes; Secondly, use height_avg and k to set the line spacing threshold y_threshold = height_avg*k, where k is an empirical coefficient; Next, traverse the text boxes of all characters, and calculate whether the absolute values ​​of the differences between the upper and lower boundaries of all non-keyword text boxes and a certain keyword text box are less than or equal to y_threshold; if so, the non-keyword text box is a suspected keyword text box; Finally, the expanded keyword text box is obtained by using the text box with the minimum value of the left boundary x1 in the suspected keyword text box and the maximum value of the right boundary x2 in the suspected keyword text box.

3. The method for identifying tags of retired photovoltaic panels according to claim 2, characterized in that: Preset the field dependency tree for keywords and their adjacent fields.

4. The method for identifying tags of retired photovoltaic panels according to claim 1, characterized in that: The formula for the boundary value of pixel coordinates is: x_min=(x_center-width / 2)*W x_max=(x_center+width / 2)*W y_min=(y_center-height / 2)*H y_max=(y_center+heignt / 2)*H Among them, x_center, y_center are the normalized center coordinates of the keyword text box; width, height are the normalized width and height of the expanded keyword text box; W, H are the width and height of the original image respectively.

5. The method for identifying tags of retired photovoltaic panels according to claim 1, characterized in that: The text recognition model is an OCR model.

6. The method for identifying tags of retired photovoltaic panels according to claim 1, characterized in that: Network detection models include: Faster R-CNN, SSD, Mask R-CNN, and RetinaNet.

7. The method for identifying tags of retired photovoltaic panels according to claim 1, characterized in that: LabelImg is used to label photovoltaic panel images.

8. The method for identifying tags of retired photovoltaic panels according to claim 1, characterized in that: Data augmentation preprocessing is performed on photovoltaic panel images.

9. A label identification system for retired photovoltaic panels, characterized in that: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the method for identifying tags of retired photovoltaic panels as described in any one of claims 1 to 8.

10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the method for identifying tags of retired photovoltaic panels as claimed in any one of claims 1 to 8 is implemented.

Images