Telegraph pole nameplate identification method and system based on improved YOLOv11
By improving the combination of YOLOv11 and CSAF modules, the problems of segmentation error and difficulty in vertical text recognition in utility pole nameplate recognition were solved, achieving high accuracy and stability in nameplate information extraction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG TONGGUANG ELECTRONICS CO LTD
- Filing Date
- 2025-12-06
- Publication Date
- 2026-04-14
AI Technical Summary
Existing methods for identifying utility pole nameplates suffer from problems such as low information extraction completeness and accuracy due to segmentation effects, especially difficulty in recognizing vertical text, and insufficient stability of recognition results.
An improved version of YOLOv11 was used to detect the area, text lines, and characters of utility pole nameplates. A CSAF module was introduced to improve the localization accuracy of small targets. Vertical text was reconstructed by combining a direction determination model. The accuracy and stability of recognition were improved by verifying the recognition results of multiple frames and using the minimum edit distance algorithm.
It significantly improves the accuracy of nameplate text detection and recognition in complex scenarios, overcomes the problem of misjudgment of vertical text, and improves the completeness and stability of recognition results.
Smart Images
Figure CN121861671A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for recognizing utility pole nameplates based on an improved YOLOv11. Background Technology
[0002] In power system operation and maintenance and asset management, the automatic identification of pole nameplate information is of great significance for realizing digital equipment management and improving inspection efficiency. Traditional manual data entry methods are inefficient and prone to errors, and can no longer meet the needs of intelligent management of large-scale power grid facilities. In recent years, with the development of computer vision and deep learning technologies, image-based automatic text detection and recognition technologies have been widely used in industrial scenarios. This includes using devices such as drones and handheld terminals to collect images of pole nameplates and using OCR (Optical Character Recognition) technology to automatically extract nameplate information.
[0003] Current mainstream text detection methods, such as EAST (Efficient and Accurate Scene Text Detector) and DBNet (Differentiable Binarization Network), rely on segmentation accuracy for their results. Using these methods, complete text lines can be segmented into multiple independent text blocks or even single characters, making it difficult to form a complete semantic structure in subsequent recognition stages and affecting the completeness and accuracy of information extraction. Furthermore, their design is primarily geared towards horizontally arranged text lines, with limited ability to detect vertically oriented or irregularly arranged text. However, in actual utility pole nameplate images, there is often vertically written text (such as "High Voltage Danger" or "Line Name").
[0004] Furthermore, existing OCR systems generally lack effective post-processing mechanisms for recognition results, leading to decreased stability and insufficient robustness of the recognition results.
[0005] In conclusion, there is an urgent need for an end-to-end identification method optimized for the characteristics of utility pole nameplates to solve the above problems. Summary of the Invention
[0006] The purpose of this invention is to provide a method for identifying utility pole nameplates based on an improved YOLOv11, which solves the problems of low information extraction completeness and accuracy caused by the segmentation effect of existing nameplate recognition methods, as well as the difficulties in recognizing vertical text and the lack of stability of the final recognition results.
[0007] This invention provides a method for identifying utility pole nameplates based on an improved YOLOv11, comprising the following steps: The improved YOLOv11 is used to detect the area of the utility pole nameplate and the text lines and characters within the area of the utility pole nameplate in the video frame, and outputs the area bounding box of the utility pole nameplate, the text line bounding box, and the character bounding box; Based on the text line box and character box, crop to obtain the text line image and character image, and determine the direction of the text line; If the text is vertical, it is converted into horizontal text before being input into the text recognition model to obtain preliminary string recognition results. The initial string recognition results are stored. When the initial string recognition results of N consecutive frames of the same utility pole nameplate are stored, the initial string recognition results are verified and the final string recognition result is output.
[0008] In some embodiments, the improved YOLOv11 Neck network introduces a CSAF (Cross-Scale Attention Fusion) module based on an attention mechanism. The fused feature map output by the CSAF module is refined by the C3K2 module and then fed into the small target detection branch of the Head network. The CSAF module generates a spatial attention weight map based on the shallow feature map through an attention mechanism, and multiplies the weight map element-wise with the upsampled deep feature map and then adds it to the shallow feature map to obtain the fused feature map. This achieves cross-scale fusion with spatial detail preservation and background noise suppression, thereby improving the localization accuracy of small targets.
[0009] In some embodiments, generating a spatial attention weight map based on a shallow feature map using an attention mechanism specifically includes: For shallow feature maps Each pixel in Calculate the square of the difference between it and all pixels in its neighborhood, sum them, and normalize: ,in, Represents pixels The neighborhood of N is the number of pixels in the neighborhood. based on Generate attention weights : ,in, yes A certain form of normalization, where k is a very small positive constant used to prevent the dividend from being 0. In some embodiments, determining the direction of the text line specifically includes: The column pixel sum and row pixel sum of the text image are statistically analyzed, and their distribution dispersion is measured by Shannon entropy. A first direction determination factor is constructed based on the degree of dispersion. The second direction determination factor is obtained by comparing the aspect ratio of the text image with an empirical threshold. , ,in, , is an empirical threshold, which is confirmed according to empirical values; After mapping the first direction determination factor to [-1, 1] using a non-linear activation function, it is linearly added to the second direction determination factor according to weights to obtain a direction determination model , and the direction determination rule is that if is less than 0, it is vertical text, otherwise it is horizontal text, , where, is the first direction determination factor after non-linear mapping, .
[0010] In some embodiments, converting vertical text into horizontal text specifically includes: For each vertical text line box, traverse all character boxes , calculate the intersection ratio, if it is higher than the ratio threshold, it is determined that the character box belongs to the vertical text line box, and the character images of the same vertical text box are sorted in ascending order according to their center point Y coordinates to form a character sequence from top to bottom; Based on the maximum height h_max in the character sequence, for any character box with height h < h_max, fill ⌊Δ / 2⌋ rows of target pixels on both the upper and lower sides, where Δ = h_max - h; if Δ is odd, one more row of target pixels is filled on the lower side than on the upper side to make the filled box height exactly equal to h_max, and the target pixel is the average value of all pixels on the four boundaries of the corresponding character box, namely the upper, lower, left, and right; The character images with unified height are horizontally spliced in the order of the above character sequence to reconstruct a horizontal text image.
[0011] In some embodiments, verifying the preliminary string recognition result specifically includes: In the preliminary string recognition results of consecutive N frames, count each type of non-repeated string and its quantity, and record it as string set S. If the category is 1, then this string is the final string recognition result; Otherwise, calculate the similarity between each string in set S and all inventory strings in the pole nameplate text database based on the minimum edit distance algorithm. For each string in set S, obtain the inventory string with the highest similarity as its corresponding standard string, and count the number of votes obtained for each type of standard string, and record it as string set D; Count the string categories in string set D. If the number of categories is 1, the strings in string set D are the final string recognition results. Otherwise, the standard string with the highest number of votes is the final string recognition result.
[0012] In some embodiments, the present invention further includes: Set the status flag sendFlag and the invalid frame counter inValidCount. Initially, sendFlag=false and inValidCount=0. In response to outputting and reporting the final string recognition result, the stored preliminary string recognition result is cleared and sendFlag is set to true. When sendFlag is true, even if the nameplate is recognized, the preliminary string recognition result will no longer be stored. In response to the nameplate recognition result changing from present to absent, inValidCount starts to accumulate the number of frames in which no nameplate was detected. When the count reaches M, sendFlag is set to false. M is set according to the frame rate and the moving speed of the image acquisition device. In response to the nameplate recognition result changing from zero to positive, inValidCount is reset to 0, and a new round of nameplate detection and recognition begins.
[0013] In another aspect, the present invention also provides a utility pole nameplate recognition system based on an improved YOLOv11, comprising: Text line and character detection module: Uses the improved YOLOv11 to detect the pole nameplate area and the text lines and characters within the pole nameplate area in the video frame, and outputs the pole nameplate area box, text line box, and character box; Text line direction determination module: Based on the text line box and character box, crop and obtain the text line image and character image, and determine the direction of the text line; Vertical text reconstruction module: If the text is vertical, it is converted into horizontal text before being input into the text recognition model to obtain preliminary string recognition results; Post-processing module for recognition results: Stores the preliminary string recognition results. When the preliminary string recognition results of N consecutive frames of the same utility pole nameplate are stored, the preliminary string recognition results are verified and the final string recognition result is output.
[0014] The beneficial effects of this invention are as follows: Compared to text detection methods such as EAST, this invention uses an improved YOLOv11 to achieve direct regression of multi-granularity target boxes at the character level, text line level, and nameplate level. This avoids the information loss and reduced recognition integrity caused by segmentation errors in traditional methods, and significantly improves the detection and recognition accuracy of nameplate text in complex scenarios. The orientation determination model of this invention comprehensively determines the orientation of text lines by combining the geometric information and pixel dispersion of the text image, and performs image reconstruction for vertical text, overcoming the problem of misjudging vertical text by traditional methods, and greatly improving the accuracy of orientation recognition and vertical text detection. By constructing a hierarchical and progressive post-processing error correction system, the accuracy and stability of the final output results can be improved. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 Here is a diagram of the existing YOLOv11 network architecture; Figure 2 The diagram shows the structure of the improved YOLOv11 network and the CSAF module network of this invention. Figure 3 This is a flowchart of the nameplate recognition method of the present invention. Detailed Implementation
[0017] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.
[0018] like Figure 1-3 As shown, this invention discloses a method for identifying utility pole nameplates based on an improved YOLOv11, comprising the following steps: S1. Use the improved YOLOv11 to detect the pole sign area and the text lines and characters within the pole sign area in the video frame, and output the pole sign area box, text line box, and character box; Compared to text detection methods such as EAST, this invention uses an improved YOLOv11 to achieve direct regression of multi-granularity target boxes at the character level, text line level, and nameplate level. This avoids the information loss and reduced recognition integrity caused by segmentation errors in traditional methods, and significantly improves the detection and recognition accuracy of nameplate text in complex scenarios.
[0019] The target bounding box coordinates mentioned above are all in the standard image coordinate system with the top left corner of the video frame as the origin, +x to the right, and +y down.
[0020] YOLOv11 includes three modules: backbone, neck, and head. The improved YOLOv11 of this invention is an improvement on the neck module.
[0021] In some embodiments, such as Figure 2 As shown, the improved YOLOv11 Neck network introduces a CSAF module (cross-scale attention fusion module) based on an attention mechanism. The fused feature map output by the CSAF module is refined by the C3K2 module and then fed into the small target detection branch of the Head network. The CSAF module generates a spatial attention weight map based on the shallow feature map through an attention mechanism, and multiplies the weight map element-wise with the upsampled deep feature map and then adds it to the shallow feature map to obtain the fused feature map. This achieves cross-scale fusion with spatial detail preservation and background noise suppression, thereby improving the localization accuracy of small targets.
[0022] like Figure 2 As shown, the backbone network includes 5 downsampling operations, outputting shallow features P3, medium features P4, and deep features P5.
[0023] Shallow feature P3 is defined as The deep feature P5 is upsampled and then concatenated with the mid-level feature P4 to form the deep feature. Shallow feature map The resolution is 1 / 8 of the input image, and the number of channels is 512; deep features The resolution is 1 / 32 of the input image, and the number of channels is 1024. yes Features after upsampling.
[0024] Shallow features A spatial attention weight map is generated through an attention mechanism, specifically including: For shallow feature maps Each pixel in Calculate the square of the difference between it and all pixels in its neighborhood, sum them, and normalize: ,in, Represents pixels The neighborhood of N is the number of pixels in the neighborhood. based on Generate attention weights : ,in, yes A certain form of normalization, where k is a very small positive constant used to prevent the dividend from being 0.
[0025] The generated attention weight map , and the upsampled deep feature map Element-wise multiplication yields a weighted feature map. .
[0026] Generate fusion features:
[0027] splicing and fusion features and, , Features It is fed into the small target detection head.
[0028] This invention incorporates a CSAF module based on an attention mechanism. This structural optimization brings the dual benefits of enhanced detail and semantic focus: on the one hand, spatial attention weights strengthen key regions in the image, significantly improving the detection capability of small-sized text in utility pole signs; on the other hand, through cross-scale feature fusion, deep semantic information and shallow detail features complement each other, enabling the model to accurately focus on text regions even in complex backgrounds, thereby comprehensively improving detection accuracy and robustness. Furthermore, the introduction of the CSAF module eliminates the need to modify downstream structures.
[0029] S2. Based on the text line frame and character frame, crop to obtain the text line image and character image, and determine the direction of the text line; In some embodiments, determining the direction of the text line specifically includes: S201. Statistically calculate the column pixel sum and row pixel sum of the text image, and use Shannon entropy to measure the dispersion of its distribution. Construct a first direction determination factor based on the degree of dispersion. Let the text line region image be... The sum of the pixels in its j-th column is The sum of the pixels in the i-th row is: ; The sum of pixels in each column and the sum of pixels in each row are used to form column sequences and row sequences, respectively. The degree of dispersion of the column sequences and row sequences is calculated based on Shannon entropy. , , It represents the degree of dispersion of the column sequence. It represents the degree of dispersion of the row sequence. It is the sum of all pixels in the text line image.
[0030] Construct the first direction decision factor:
[0031] , The value is taken as a very small positive number to prevent The value is 0, for example, it can be taken as 0. .
[0032] S202. Based on the comparison between the aspect ratio of the text image and an empirical threshold, obtain the second direction determination factor. , ,in, , This is an empirical threshold, defined based on experience, for example, =1.5 S203. After mapping the first direction determination factor to [-1,1] using a nonlinear activation function, the second direction determination factor is linearly added according to its weights to obtain the direction determination model. The direction determination rule is: if A value less than 0 indicates vertical text; otherwise, it indicates horizontal text. Using a nonlinear activation function The first direction determination factor Mapped to [-1, 1] In the formula, α is set to 1.0 by default and is a scaling factor that controls the response sensitivity; In the formula, , The specific value of λ is usually determined using a validation set. A series of λ values (such as 0, 0.1, 0.2, ..., 0.9, 1.0) are systematically tried using methods such as grid search, and then the value that satisfies the λ value is selected. The λ value that achieves the highest overall accuracy on the validation set is used as the final parameter.
[0033] The orientation determination model of this invention comprehensively determines the orientation of text lines by combining the geometric information and pixel dispersion of the text image, overcoming the problem of misjudging vertical text by traditional methods and greatly improving the orientation recognition accuracy.
[0034] S3. If the text is vertical, convert it to horizontal text and then input it into the text recognition model to obtain preliminary string recognition results. In some embodiments, converting vertical text into horizontal text specifically includes: S301. For each vertical text line box, iterate through all character boxes. Calculate the intersection ratio. If it is higher than the ratio threshold, the character box is determined to belong to the vertical text line box. Sort the character images of the same vertical text box according to the Y coordinate of its center point from smallest to largest to form a top-down character sequence. The ratio threshold is generally set to 70%.
[0035] S302. Taking the maximum height h_max in the character sequence as a benchmark, for any character box with a height h < h_max, fill ⌊Δ / 2⌋ rows of target pixels on both the upper and lower sides, where Δ = h_max - h; if Δ is odd, then one more row of target pixels is filled on the lower side than on the upper side so that the height of the filled box is exactly equal to h_max. The target pixels are the average value of all the pixels on the upper, lower, left, and right four boundaries of the corresponding character box. S303. Horizontally splice the character images with unified height in the order of the above character sequence to reconstruct a horizontal text image.
[0036] If it is horizontal text, it is directly input into the text recognition model. If it is vertical text, it needs to be reconstructed into a horizontal text image according to the above steps.
[0037] According to the character box annotation rule, the edge pixel values are already the background color of the nameplate. Filling the height deficiency with the average value of its edge pixels maintains the consistency of structure and vision. Finally, the adjusted character images are horizontally spliced in order to reconstruct a new image and sent to the character recognition module. This method effectively solves the problem of mis-segmentation of vertical text in traditional text detection. The reconstructed image can be directly input into the standard OCR recognition model without modifying the network structure of OCR recognition.
[0038] S4. Store the preliminary string recognition results. When the preliminary string recognition results of the same utility pole nameplate for N consecutive frames are stored, verify the preliminary string recognition results and output the final string recognition result.
[0039] In some embodiments, verifying the preliminary string recognition results specifically includes: S401. In the preliminary string recognition results of N consecutive frames, count each type of non-repeated string and its quantity, and record it as the string set S. If the category of non-repeated strings is 1, then this string is the final string recognition result. For example, if N is set to 15, and the string set S has non-repeated strings "1okV Wuhe Line" and "10kV Wuhe Qian" with quantities of 5 and 10 respectively, then the category of non-repeated strings is 2.
[0040] The application scenario of the present invention is to perform real-time inspection and nameplate recognition on utility pole nameplates. The selection of N needs to be based on the frame rate and the moving speed of the image acquisition device to ensure that more than N frames of images are recorded for the same utility pole nameplate.
[0041] S402. If the non-repetitive string category is not 1, calculate the similarity between each string in set S and all the inventory strings in the pole nameplate text database based on the minimum edit distance algorithm. For each string in set S, obtain the inventory string with the highest similarity as its corresponding standard string, and count the number of votes for each type of standard string, which is denoted as string set D. The minimum edit distance algorithm refers to the minimum number of single-character edit operations required to convert one string into another string. The allowed operations are inserting, deleting, and replacing characters. For example, for the string "1okV Wuhe Line" in set S with a quantity of 5 and the string "10kV Wuhe Qian" in set S with a quantity of 10, based on the calculation of the minimum edit distance algorithm, both have the highest similarity with the inventory string "10kV Wuhe Line". Then, the inventory string "10kV Wuhe Line" is the standard string for the strings "1okV Wuhe Line" and "10kV Wuhe Qian" in set S, and the number of votes for this standard string "10kV Wuhe Line" is 15.
[0042] S403. Count the string categories in string set D. If the number of categories is 1, the strings in string set D are the final string recognition results. Otherwise, the standard string with the highest number of votes is the final string recognition result. The minimum edit distance algorithm is a relatively well-known algorithm. This article will not elaborate on it in detail but only give a brief overview as follows. Let be the edit distance between strings s and t. After obtaining based on the minimum edit distance algorithm, perform normalization processing . , is the string length. Here, when , . When the two strings are completely unrelated and very different, , represents the similarity between strings s and t.
[0043] The present invention establishes and maintains a buffer for historical recognition results, and combines with the nameplate text database to establish a hierarchical and progressive post-processing error correction system. This solution brings multiple beneficial effects: based on multiple-frame recognition results and performing repetitive filtering, it effectively eliminates instantaneous misidentifications in the video stream scenario; the similarity calculation based on the minimum edit distance algorithm can fully utilize the database text to correct similar and easily confused string errors; and the voting mechanism improves the accuracy and stability of the final output result by comprehensively considering historical recognition results and database matching.
[0044] In some embodiments, the present invention further includes: Set the status flag sendFlag and the invalid frame counter inValidCount. Initially, sendFlag=false and inValidCount=0. In response to outputting and reporting the final string recognition result, the stored preliminary string recognition result is cleared and sendFlag is set to true. When sendFlag is true, even if the nameplate is recognized, the preliminary string recognition result will no longer be stored. In response to the nameplate recognition result changing from present to absent, inValidCount starts to accumulate the number of frames in which no nameplate was detected. When the count reaches M, sendFlag is set to false. M is set according to the frame rate and the moving speed of the image acquisition device. In response to the nameplate recognition result changing from zero to positive, inValidCount is reset to 0, and a new round of nameplate detection and recognition begins.
[0045] This invention sets a status flag bit sendFlag to record whether the final string recognition result has been reported, and sets an invalid frame counter inValidCount to record the number of frames in which the nameplate was not detected, to prevent the same nameplate information from being reported repeatedly and to prevent confusion between different pole nameplates. The initial state is sendFlag=false and inValidCount=0. When a utility pole nameplate is detected, the initial string recognition results of N consecutive frames are stored. Then, the initial string recognition results are verified. After the final string recognition results are output and reported, the stored string recognition results are cleared and sendFlag is set to true. When sendFlag is true, even if the nameplate is recognized, the initial string recognition result will no longer be stored, and the current pole nameplate recognition is complete. The application scenario of this invention is real-time inspection and recognition of utility pole nameplates. In actual scenarios, there will be a certain distance between two utility poles. Therefore, as the image acquisition device moves to the next utility pole, there will be a period of time when no nameplate is detected, and the nameplate detection result will change from present to absent. When no nameplate is detected, inValidCount starts to accumulate. When it accumulates to M (default is M = frame rate * 6s), sendFlag is set to false, waiting for a new utility pole nameplate to be detected. Then, inValidCount is reset to 0, and a new round of nameplate detection and recognition begins.
[0046] Another aspect of the present invention discloses a utility pole nameplate recognition system based on an improved YOLOv11, comprising: Text line and character detection module: Uses the improved YOLOv11 to detect the pole nameplate area and the text lines and characters within the pole nameplate area in the video frame, and outputs the pole nameplate area box, text line box, and character box; Text line direction determination module: Based on the text line box and character box, crop and obtain the text line image and character image, and determine the direction of the text line; Vertical text reconstruction module: If the text is vertical, it is converted into horizontal text before being input into the text recognition model to obtain preliminary string recognition results; Post-processing module for recognition results: Stores the preliminary string recognition results. When the preliminary string recognition results of N consecutive frames of the same utility pole nameplate are stored, the preliminary string recognition results are verified and the final string recognition result is output.
[0047] The above provides a detailed description of the method and system for identifying utility pole nameplates based on an improved YOLOv11, as provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of these embodiments are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.
[0048] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
Claims
1. A method for identifying utility pole nameplates based on an improved YOLOv11, characterized in that, Includes the following steps: The improved YOLOv11 is used to detect the area of the utility pole nameplate and the text lines and characters within the area of the utility pole nameplate in the video frame, and outputs the area bounding box of the utility pole nameplate, the text line bounding box, and the character bounding box; Based on the text line box and character box, crop to obtain the text line image and character image, and determine the direction of the text line; If the text is vertical, it is converted into horizontal text before being input into the text recognition model to obtain preliminary string recognition results. The initial string recognition results are stored. When the initial string recognition results of N consecutive frames of the same utility pole nameplate are stored, the initial string recognition results are verified and the final string recognition result is output.
2. The method for identifying utility pole nameplates based on improved YOLOv11 according to claim 1, characterized in that, The improved YOLOv11 Neck network introduces a CSAF (Cross-Scale Attention Fusion) module based on an attention mechanism. The fused feature map output by the CSAF module is refined by the C3K2 module and then fed into the small target detection branch of the Head network. The CSAF module generates a spatial attention weight map based on the shallow feature map through an attention mechanism. This weight map is then multiplied element-wise with the upsampled deep feature map and added back to the shallow feature map to obtain the fused feature map. This achieves cross-scale fusion with spatial detail preservation and background noise suppression, thereby improving the localization accuracy of small targets.
3. The method for identifying utility pole nameplates based on the improved YOLOv11 according to claim 2, characterized in that, The specific steps of generating a spatial attention weight map based on shallow feature maps using an attention mechanism include: For shallow feature maps Each pixel in Calculate the square of the difference between it and all pixels in its neighborhood, sum them, and normalize: ,in, Represents pixels The neighborhood of N is the number of pixels in the neighborhood. based on Generate attention weights : ,in, yes A certain form of normalization, where k is a very small positive constant used to prevent the dividend from being 0.
4. The method for identifying utility pole nameplates based on improved YOLOv11 according to claim 1, characterized in that, The determination of the direction of the text line specifically includes: The column pixel sum and row pixel sum of the text image are statistically analyzed, and their distribution dispersion is measured by Shannon entropy. A first direction determination factor is constructed based on the degree of dispersion. The second direction determination factor is obtained by comparing the aspect ratio of the text image with an empirical threshold. , ,in, , This is an empirical threshold, confirmed based on empirical values; After mapping the first direction determination factor to [-1, 1] using a nonlinear activation function, the second direction determination factor is linearly added according to its weights to obtain the direction determination model. The direction determination rule is: if A value less than 0 indicates vertical text; otherwise, it indicates horizontal text. In the formula, It is the first direction determination factor after nonlinear mapping. .
5. The method for identifying utility pole nameplates based on improved YOLOv11 according to claim 1, characterized in that, Converting vertical text to horizontal text specifically includes: For each vertical text line box, iterate through all character boxes. Calculate the intersection ratio. If it is higher than the ratio threshold, the character box is determined to belong to the vertical text line box. Sort the character images of the same vertical text box according to the Y coordinate of its center point from small to large to form a character sequence from top to bottom. Based on the maximum height h_max in the character sequence, for any character box with height h < h_max, fill the top and bottom sides with ⌊Δ / 2⌋ rows of target pixels, where Δ = h_max − h; if Δ is odd, the bottom side is padded with 1 more row of target pixels than the top side, so that the height of the box after filling is exactly equal to h_max. The target pixels are the average of all pixels on the top, bottom, left, and right boundaries of the corresponding character box. The character images with uniform height are horizontally stitched together according to the above character sequence to reconstruct a horizontal text image.
6. The method for identifying utility pole nameplates based on improved YOLOv11 according to claim 1, characterized in that, The preliminary string recognition results are validated, specifically including: In the preliminary string recognition results of N consecutive frames, count the number of non-repeating strings in each class and record them as string set S. If the class is 1, then the string is the final string recognition result. Otherwise, the similarity between each string in set S and all stock strings in the telephone pole nameplate text database is calculated based on the minimum edit distance algorithm. For each string in set S, the stock string with the highest similarity is taken as its corresponding standard string, and the number of votes for each type of standard string is counted, denoted as string set D. Count the number of string categories in string set D. If the number of categories is 1, the string in string set D is the final string recognition result; otherwise, the standard string with the highest number of votes is the final string recognition result.
7. The method for identifying utility pole nameplates based on improved YOLOv11 according to claim 1, characterized in that, Also includes: Set the status flag sendFlag and the invalid frame counter inValidCount. Initially, sendFlag=false and inValidCount=0. In response to outputting and reporting the final string recognition result, the stored preliminary string recognition result is cleared and sendFlag is set to true. When sendFlag is true, even if the nameplate is recognized, the preliminary string recognition result will no longer be stored. In response to the nameplate recognition result changing from present to absent, inValidCount starts to accumulate the number of frames in which no nameplate was detected. When the count reaches M, sendFlag is set to false. M is set according to the frame rate and the moving speed of the image acquisition device. In response to the nameplate recognition result changing from zero to positive, inValidCount is reset to 0, and a new round of nameplate detection and recognition begins.
8. A utility pole nameplate recognition system based on an improved YOLOv11, characterized in that, include: Text line and character detection module: Uses the improved YOLOv11 to detect the pole nameplate area and the text lines and characters within the pole nameplate area in the video frame, and outputs the pole nameplate area box, text line box, and character box; Text line direction determination module: Based on the text line box and character box, crop and obtain the text line image and character image, and determine the direction of the text line; Vertical text reconstruction module: If the text is vertical, it is converted into horizontal text before being input into the text recognition model to obtain preliminary string recognition results; Post-processing module for recognition results: Stores the preliminary string recognition results. When the preliminary string recognition results of N consecutive frames of the same utility pole nameplate are stored, the preliminary string recognition results are verified and the final string recognition result is output.