A Method and System for Identifying and Locating the Nameplate Area in a Nameplate Image

Abstract

The present invention discloses a method and system for identifying and positioning a nameplate area in a nameplate image, which relates to the technical field of image processing; the first image is subjected to exposure adjustment to obtain a second image, the contours in the second image are extracted, and a set of areas to be detected is obtained according to the extracted contours; for the set of areas to be detected, the second image is cut with the area to be detected to obtain a detection image, the text contours of the detection image are extracted to obtain a target text layout, and the target text layout is substituted into a preset database for similarity comparison. If the comparison result shows the existence of a nameplate image, a perspective transformation matrix is performed on the detection image to obtain a final image. The preprocessing and exposure adjustment of the initial image reduce the problems of overexposure or underexposure of the image caused by environmental light, improve the recognition accuracy of the nameplate information, and perform contour extraction on the second image and cutting of the set of areas to be detected, so as to accurately locate the nameplate area and improve the accuracy and efficiency of nameplate recognition and positioning.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method and system for identifying and positioning a nameplate area in a nameplate image. Background Art

[0002] The scale of power infrastructure is huge, with a large number of equipment, widely distributed, constituting a complex and huge power system network. In order to effectively manage these devices and ensure the stable operation of the power system, it is usually necessary to record key device information, such as model, specification, production date, manufacturer, etc., in detail on the nameplate. However, at present, the information registration work of electrical equipment nameplates mainly relies on manual operation, which is not only inefficient, consuming a large amount of manpower and time, but also prone to errors, such as information entry errors, omissions or repetitions. At the same time, this traditional manual method will also cause serious loss of human resources, which is not conducive to the long-term development and intelligent transformation of the power system.

[0003] To solve the above problems, currently, the image of the electrical equipment nameplate is usually obtained by photographing the equipment, and then the information on the nameplate is extracted through image analysis technology. This method improves the efficiency and accuracy of information registration to a certain extent, but there are still some challenges. In practical applications, since some electrical equipment in the power system needs to be installed outdoors, and due to the high installation position or environmental restrictions, it is often impossible to directly photograph the nameplate on the electrical equipment from the front. This situation leads to overexposure or underexposure of the image due to angle problems during the process of collecting the nameplate image of the electrical equipment, making it very difficult to identify the nameplate subsequently, and even resulting in the inability to accurately identify the nameplate and the key information on the nameplate, thus reducing the accuracy and efficiency of nameplate identification and positioning. Summary of the Invention

[0004] The object of the present invention is to solve the problem that the image is prone to overexposure or underexposure due to angle problems, which makes it very difficult to identify the nameplate subsequently, and even leads to the inability to accurately identify the nameplate and the key information on the nameplate, thus reducing the accuracy and efficiency of nameplate identification and positioning, and to propose a method and system for identifying and positioning a nameplate area in a nameplate image.

[0005] In the first aspect of the implementation of the present invention, first, a method for identifying and positioning a nameplate area in a nameplate image is proposed. The method includes:

[0006] Obtain an initial image, and preprocess the initial image to obtain a first image; the initial image is an image containing a nameplate area;

[0007] Adjust the exposure of the first image to obtain a second image, extract the contours in the second image, and obtain a set of regions to be detected according to the extracted contours;

[0008] For the set of regions to be detected, cut the second image with the regions to be detected to obtain a detection image, extract the text contour of the detection image to obtain the target text layout, substitute the target text layout into a preset database for similarity comparison. If the comparison result indicates the existence of a nameplate image, mark the detection image as the target image;

[0009] Perform a perspective transformation matrix on the target image to obtain a final image.

[0010] Optionally, performing exposure adjustment on the first image to obtain the second image includes:

[0011] Process the first image through a tone mapping curve to obtain a set of exposure images;

[0012] For each exposure image in the set of exposure images, perform Laplacian filtering on the exposure image to obtain the filtering result value of each pixel in the exposure image, record the absolute value of the filtering result value as the first parameter value, and obtain a first parameter map based on the first parameter values of all pixels;

[0013] For each pixel of the exposure image, calculate the mean value of the pixel in the R channel, G channel, and B channel, obtain the standard deviation corresponding to the pixel based on the mean value, and obtain a second parameter map based on the standard deviations corresponding to all pixels;

[0014] For each pixel of the exposure image, calculate the brightness value corresponding to the pixel, obtain the exposure value corresponding to the pixel based on the brightness value, and obtain a third parameter map based on the exposure values corresponding to all pixels;

[0015] Obtain the weight value corresponding to the exposure image based on the first parameter map, second parameter map, and third parameter map corresponding to the exposure image, and perform weighted summation on all exposure images based on the weight values corresponding to all exposure images to obtain the second image.

[0016] Optionally, extracting the contours in the second image and obtaining the set of regions to be detected based on the extracted contours includes:

[0017] Substitute the second image into a deep aggregation network to obtain a first multi-scale feature map, and perform center point, offset, and direction prediction on the first multi-scale feature map to obtain a plurality of predicted bounding boxes;

[0018] For each predicted bounding box, perform corner detection on the predicted bounding box to obtain a set of target corner points, and connect all the points in the set of target corner points to obtain an initial connection line set;

[0019] For the initial connection line set, obtain the initial connection line combination with the largest formed area to obtain a first connection line set;

[0020] Bring the first connection line set and the second image into the connection line enhancement model to obtain a second connection line set. Form a region to be detected according to the connection lines in the second connection line set, and obtain a set of regions to be detected according to the regions to be detected corresponding to each predicted bounding box.

[0021] Optionally, bringing the first connection line set and the second image into the connection line enhancement model to obtain a second connection line set includes:

[0022] Substitute the second image into the CNN backbone network to obtain a second multi-scale feature map, and encode and optimize the second multi-scale feature map through a transformer encoder to obtain a target-scale feature map;

[0023] For each connection line in the first connection line set, perform position encoding on the endpoint coordinates of the connection line to obtain a first encoded value, and substitute the first encoded value into a multi-layer perceptron to obtain a target query vector;

[0024] Bring the target query vector and the target-scale feature map into a transformer decoder for decoding to obtain a target enhanced connection line, enhance the connection line according to the target enhanced connection line to obtain a target connection line, and obtain the target connection lines corresponding to all the connection lines in the first connection line set to obtain a second connection line set.

[0025] Optionally, performing text contour extraction on the detection image to obtain a target text layout includes:

[0026] Convert the detection image into a grayscale image, and convert the grayscale image into a binary image through an edge detection algorithm to obtain a target detection image;

[0027] Perform Hough transform detection on the target detection image to obtain the boundary of the text region in the target detection image, and determine the text recognition contour according to the text region boundary;

[0028] Perform text cutting on the target detection image according to the text recognition contour to obtain a target text image;

[0029] Obtain all target text images for text recognition to obtain a character label set, and layout all the character labels in the character label set according to the recognition positions of all the target text images to obtain a target text layout.

[0030] In the second aspect of the implementation of the present invention, a system for identifying and positioning a nameplate area in a nameplate image is proposed, including:

[0031] A first image acquisition module, configured to acquire an initial image, and preprocess the initial image to obtain a first image; the initial image is an image containing a nameplate area;

[0032] The contour acquisition module is used to perform exposure adjustment on the first image to obtain a second image, extract the contours in the second image, and obtain a set of regions to be detected according to the extracted contours;

[0033] The target image acquisition module is used to, for the set of regions to be detected, cut the second image with the regions to be detected to obtain a detection image, perform text contour extraction on the detection image to obtain a target text layout, substitute the target text layout into a preset database for similarity comparison. If the comparison result indicates the existence of a nameplate image, the detection image is recorded as the target image;

[0034] The final image acquisition module is used to perform a perspective transformation matrix on the target image to obtain a final image.

[0035] Optionally, the contour acquisition module includes:

[0036] The exposure image generation module is used to process the first image through a tone mapping curve to obtain a set of exposure images;

[0037] The first parameter map acquisition module is used to, for each exposure image in the set of exposure images, perform Laplacian filtering on the exposure image to obtain the filtering result value of each pixel in the exposure image, record the absolute value of the filtering result value as the first parameter value, and obtain a first parameter map according to the first parameter values of all pixels;

[0038] The second parameter map acquisition module is used to, for each pixel of the exposure image, calculate the mean value of the pixel on the R channel, G channel, and B channel, obtain the standard deviation corresponding to the pixel according to the mean value, and obtain a second parameter map according to the standard deviations corresponding to all pixels;

[0039] The third parameter map acquisition module is used to, for each pixel of the exposure image, calculate the brightness value corresponding to the pixel, obtain the exposure value corresponding to the pixel according to the brightness value, and obtain a third parameter map according to the exposure values corresponding to all pixels;

[0040] The second image generation module is used to obtain the weight value corresponding to the exposure image according to the first parameter map, second parameter map, and third parameter map corresponding to the exposure image, and perform weighted summation on all exposure images according to the weight values corresponding to all exposure images to obtain a second image.

[0041] Optionally, the contour acquisition module further includes:

[0042] The predicted bounding box generation module is used to substitute the second image into a deep aggregation network to obtain a first multi-scale feature map, and perform center point, offset, and direction prediction on the first multi-scale feature map to obtain a plurality of predicted bounding boxes;

[0043] An initial connection line generation module, which is used for each predicted bounding box, to detect corner points of the predicted bounding box to obtain a set of target corner points, and connect all points in the set of target corner points to obtain an initial connection line set;

[0044] A first connection line determination module, which is used for the initial connection line set to obtain a combination of initial connection lines with the largest formed area to obtain a first connection line set;

[0045] A region to be detected determination module, which is used to bring the first connection line set and the second image into a connection line enhancement model to obtain a second connection line set, form a region to be detected according to the connection lines in the second connection line set, and obtain a set of regions to be detected according to the regions to be detected corresponding to each predicted bounding box.

[0046] Optionally, the region to be detected determination module includes:

[0047] An encoding optimization module, which is used to bring the second image into a CNN backbone network to obtain a second multi-scale feature map, and perform encoding optimization on the second multi-scale feature map through a transformer encoder to obtain a target-scale feature map;

[0048] A target query vector determination module, which is used for each connection line in the first connection line set to perform position encoding on the endpoint coordinates of the connection line to obtain a first encoded value, and substitute the first encoded value into a multi-layer perceptron to obtain a target query vector;

[0049] A second connection line determination module, which is used to bring the target query vector and the target-scale feature map into a transformer decoder for decoding to obtain a target enhanced connection line, enhance the connection line according to the target enhanced connection line to obtain a target connection line, and obtain a second connection line set for all connection lines corresponding to the first connection line set.

[0050] Optionally, the target image acquisition module includes:

[0051] A target detection image determination module, which is used to convert the detection image into a grayscale image, and convert the grayscale image into a binary image through an edge detection algorithm to obtain a target detection image;

[0052] A text region boundary determination module, which is used to perform Hough transform detection on the target detection image to obtain the text region boundary in the target detection image, and determine the text recognition contour according to the text region boundary;

[0053] An image cutting module, which is used to cut the target detection image according to the text recognition contour to obtain a target text image;

[0054] An image layout module, which is used to obtain all target text images, perform text recognition to obtain a character tag set, and layout all character tags in the character tag set according to the recognition positions of all target text images to obtain a target text layout.

[0055] Advantages of the present invention:

[0056] The present invention provides a method for identifying and positioning a nameplate area in a nameplate image. An initial image is obtained, and the initial image is preprocessed to obtain a first image; the first image is adjusted for exposure to obtain a second image, the contours in the second image are extracted, and a set of areas to be detected is obtained according to the extracted contours; for the set of areas to be detected, the second image is cut with the area to be detected to obtain a detection image, the text contours of the detection image are extracted to obtain a target text layout, and the target text layout is substituted into a preset database for similarity comparison. If the comparison result indicates the existence of a nameplate image, the detection image is recorded as a target image; a perspective transformation matrix is performed on the target image to obtain a final image. Through the preprocessing and exposure adjustment of the initial image, the problems of overexposure or underexposure of the image caused by environmental light can be effectively reduced, thereby improving the recognition accuracy of nameplate information. Extracting the contours of the second image and cutting the set of areas to be detected helps to accurately locate the nameplate area, further improving the recognition accuracy and enhancing the accuracy and efficiency of nameplate recognition and positioning. Description of the Drawings

[0057] The present invention will be further described below with reference to the accompanying drawings.

[0058] Figure 1 It is a flowchart of a method for identifying and positioning a nameplate area in a nameplate image provided by an embodiment of the present invention;

[0059] Figure 2 It is a framework diagram of a system for identifying and positioning a nameplate area in a nameplate image provided by an embodiment of the present invention. Detailed Embodiments

[0060] Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0061] An embodiment of the present invention provides a method for identifying and positioning a nameplate area in a nameplate image. Refer to Figure 1 , Figure 1 It is a flowchart of a method for identifying and positioning a nameplate area in a nameplate image provided by an embodiment of the present invention. The method includes the following steps:

[0062] S101, obtain an initial image, and preprocess the initial image to obtain a first image.

[0063] S102. Adjust the exposure of the first image to obtain a second image, extract the contours in the second image, and obtain a set of regions to be detected based on the extracted contours;

[0064] S103. For the set of regions to be detected, cut the second image with the regions to be detected to obtain detection images, extract the text contours of the detection images to obtain the target text layout, substitute the target text layout into a preset database for similarity comparison. If the comparison result indicates the existence of a nameplate image, mark the detection image as the target image;

[0065] S104. Perform a perspective transformation matrix on the target image to obtain the final image.

[0066] Among them, the initial image is an image containing the nameplate area.

[0067] Based on a method for identifying and positioning the nameplate area in a nameplate image provided by an embodiment of the present invention, through preprocessing and exposure adjustment of the initial image, the problems of overexposure or underexposure of the image caused by environmental light can be effectively reduced, thereby improving the recognition accuracy of nameplate information. Extracting the contours of the second image and cutting the set of regions to be detected helps to accurately locate the nameplate area, further improving the recognition accuracy and efficiency of the nameplate recognition and positioning.

[0068] In one implementation, through exposure adjustment and contour extraction, the area containing nameplate information, that is, the set of regions to be detected, can be accurately identified from the second image, which is the basis for subsequent text extraction and recognition. Precise positioning can avoid misprocessing of non-nameplate areas and improve the processing efficiency.

[0069] In one implementation, the preprocessing includes denoising, filtering, enhancement, etc., which are used to improve the image quality, enhance the image features, and simplify the image data for subsequent image analysis.

[0070] In one implementation, the preset database stores the text layouts corresponding to images of different nameplates taken at different angles. The similarity comparison can be performed through cosine similarity, Jaccard similarity, etc.

[0071] In one implementation, if after traversing the regions to be detected, all comparison results indicate the non-existence of a nameplate image, a reshooting message is sent to the photographer.

[0072] In one implementation, for each region to be detected, a detection image is obtained through cutting, and the text contours are extracted from it, which can efficiently obtain the text information on the nameplate. The processing method based on regions is more efficient than global processing and can more accurately extract the target text layout.

[0073] In one implementation, the extracted target text layout is compared with a preset database for similarity, which can intelligently determine whether the current detected image is a nameplate image. This not only improves the recognition accuracy but also has a certain degree of flexibility, capable of adapting to nameplate images of different types and formats.

[0074] In one implementation, a perspective transformation matrix process is performed on the detected image confirmed as a nameplate image, and a final image that is more in line with the human eye's viewing habit and easier for subsequent processing can be obtained. Perspective transformation can correct image distortion caused by factors such as shooting angle and camera aberration, improving the visual effect and usability of the image.

[0075] In one embodiment, performing exposure adjustment on the first image to obtain the second image includes:

[0076] Processing the first image through a tone mapping curve to obtain an exposure image set;

[0077] For each exposure image in the exposure image set, performing Laplacian filtering on the exposure image to obtain the filtering result value of each pixel in the exposure image, recording the absolute value of the filtering result value as the first parameter value, and obtaining a first parameter map based on the first parameter values of all pixels;

[0078] For each pixel of the exposure image, calculating the mean value of the pixel on the R channel, G channel, and B channel, obtaining the standard deviation corresponding to the pixel based on the mean value, and obtaining a second parameter map based on the standard deviations corresponding to all pixels;

[0079] For each pixel of the exposure image, calculating the brightness value corresponding to the pixel, obtaining the exposure value corresponding to the pixel based on the brightness value, and obtaining a third parameter map based on the exposure values corresponding to all pixels;

[0080] Obtaining the weight value corresponding to the exposure image based on the first parameter map, second parameter map, and third parameter map corresponding to the exposure image, and performing weighted summation on all exposure images based on the weight values corresponding to all exposure images to obtain the second image.

[0081] In one implementation, by using images with different exposures, details in the bright and dark parts can be better retained, thereby improving the dynamic range of the image. By performing Laplacian filtering on the image, the edges and details of the image can be enhanced, making the image clearer.

[0082] In one implementation, performing Laplacian filtering on the exposure image is through the formula L(i,j) = Σ k，lI(i+k,j+l).K(k,l) gives the filtered result value, where L(i, j) is the filtered result value of the pixel point (i, j), K(k, l) is the Laplacian kernel, k is the radius of the kernel in the horizontal direction, l is the radius of the kernel in the vertical direction, and I(i + k, j + l) is the pixel value corresponding to the pixel point I(i + k, j + l) in the exposure image.

[0083] In one implementation, through the formula the weight value corresponding to the exposure image is obtained; where W is the weight value, A is the first parameter value, B is the standard deviation, C is the exposure value, and w 1 , w 2 and w 3 are weight exponents determined by technicians.

[0084] In one implementation, through the method of weighted summation, random noise in the image can be effectively reduced. Since the noise characteristics of different exposure images are different, the noise can be smoothed through weighted summation.

[0085] In one implementation, calculating the mean and standard deviation of the RGB channels for each pixel can better adjust the color and brightness, making the image more natural and balanced. By calculating the brightness value and exposure value of the pixel, the problem of overexposure or underexposure of the image can be avoided, and the brightness distribution of the image can be made more reasonable.

[0086] In one implementation, by combining the first parameter map (Laplacian filtering result), the second parameter map (standard deviation), and the third parameter map (exposure value), the details, contrast, and brightness of the image can be comprehensively considered, thereby obtaining a higher-quality image.

[0087] In one embodiment, the contours in the second image are extracted, and the set of regions to be detected obtained according to the extracted contours includes:

[0088] The second image is substituted into the deep aggregation network to obtain the first multi-scale feature map, and multiple prediction bounding boxes are obtained by predicting the center point, offset, and direction of the first multi-scale feature map;

[0089] For each prediction bounding box, corner detection is performed on the prediction bounding box to obtain a set of target corner points, and all points in the set of target corner points are connected to obtain an initial connection line set;

[0090] For the initial connection line set, the initial connection line combination with the largest formed area is obtained to get the first connection line set;

[0091] The first connection line set and the second image are brought into the connection line enhancement model to obtain the second connection line set. The regions to be detected are formed according to the connection lines in the second connection line set, and the set of regions to be detected is obtained according to the regions to be detected corresponding to each prediction bounding box.

[0092] In one implementation, extracting multi-scale feature maps through a deep learning network can better capture complex information in images and improve the accuracy of detection; by combining information at different scales, both small and large targets can be effectively detected, enhancing the robustness of detection.

[0093] In one implementation, by performing corner detection on the predicted bounding boxes, the four corners of the target region can be accurately determined, avoiding the inaccuracy of the bounding boxes.

[0094] In one implementation, through the calculation of the connection line set and the processing of the enhancement model, connection lines can be better extracted and optimized, thereby obtaining a more accurate region to be detected.

[0095] In one embodiment, bringing the first connection line set and the second image into the connection line enhancement model to obtain the second connection line set includes:

[0096] Substituting the second image into the CNN backbone network to obtain a second multi-scale feature map, and encoding and optimizing the second multi-scale feature map through a transformer encoder to obtain a target-scale feature map;

[0097] For each connection line in the first connection line set, performing position encoding on the endpoint coordinates of the connection line to obtain a first encoded value, and substituting the first encoded value into a multi-layer perceptron to obtain a target query vector;

[0098] Bringing the target query vector and the target-scale feature map into the transformer decoder for decoding to obtain a target-enhanced connection line, enhancing the connection line according to the target-enhanced connection line to obtain a target connection line, and obtaining the target connection lines corresponding to all the connection lines in the first connection line set to obtain the second connection line set.

[0099] In one implementation, extracting multi-scale feature maps through a CNN backbone network can effectively capture multi-level information in images, providing rich features for subsequent processing; encoding and optimizing the multi-scale feature maps further enhances the expressive power of the feature maps, making the feature maps more suitable for the extraction and enhancement tasks of connection lines.

[0100] In one implementation, performing position encoding on the connection line endpoint coordinates can preserve spatial information and improve the processing effect of the multi-layer perceptron and the transformer model on the connection lines; converting the encoded value into a target query vector through a multi-layer perceptron simplifies the feature representation of the connection lines, facilitating subsequent decoding and enhancement processing.

[0101] In one implementation, inputting the target query vector and the target-scale feature map into the transformer decoder, making full use of global information and context relationships to decode and enhance the connection lines. Through the enhanced connection lines output by the decoder, the connection lines can be more accurately represented and optimized, making the connection lines more accurate and stable.

[0102] In one implementation, through positional encoding and a target query vector, the model can adapt to connection lines of different positions and scales, and has strong adaptability.

[0103] In one embodiment, performing text contour extraction on the detected image to obtain the target text layout includes:

[0104] Converting the detected image into a grayscale image, and converting the grayscale image into a binary image through an edge detection algorithm to obtain a target detection image;

[0105] Performing Hough transform detection on the target detection image to obtain the boundary of the text region in the target detection image, and determining the text recognition contour according to the boundary of the text region;

[0106] Performing text cutting on the target detection image according to the text recognition contour to obtain a target text image;

[0107] Obtaining all target text images and performing text recognition to obtain a character label set, and arranging all character labels in the character label set according to the recognition positions of all target text images to obtain the target text layout.

[0108] In one implementation, by converting the detected image into a grayscale image and applying an edge detection algorithm, the contour of the text region can be more clearly highlighted, thereby improving the accuracy of subsequent processing; converting the grayscale image into a binary image can effectively remove background noise and enhance the contrast of the text region, making the text region more obvious.

[0109] In one implementation, by detecting the boundary of the text region through Hough transform, the text region in the image can be accurately located. Hough transform is particularly effective for line detection and can reliably find the boundary line of the text region; according to the detected boundary of the text region, the text recognition contour can be further accurately determined to ensure the accuracy of subsequent text cutting.

[0110] In one implementation, performing text cutting according to the accurately determined text recognition contour can ensure that the cut text image is complete and accurate, avoiding problems such as loss of text parts or incomplete cutting. Obtaining the cut target text image helps to improve the effect of OCR text recognition because the quality of the input image directly affects the accuracy of OCR recognition.

[0111] In one implementation, by performing text recognition on the target text image through OCR technology, the text in the image can be efficiently converted into a character label set. According to the position information of text recognition, the character labels can be reasonably arranged to reconstruct the layout structure of the original text and retain the format and structure information of the text.

[0112] An embodiment of the present invention also provides a nameplate area recognition and positioning system in a nameplate image based on the same inventive concept. Refer to Figure 2 , Figure 2 which is a framework diagram of a nameplate area recognition and positioning system provided by an embodiment of the present invention, including:

[0113] A first image acquisition module, configured to acquire an initial image and preprocess the initial image to obtain a first image; the initial image is an image containing a nameplate area;

[0114] A contour acquisition module, configured to perform exposure adjustment on the first image to obtain a second image, extract the contours in the second image, and obtain a set of regions to be detected according to the extracted contours;

[0115] A target image acquisition module, configured to cut the second image with the regions to be detected in the set of regions to be detected to obtain a detection image, extract the text contour of the detection image to obtain a target text layout, substitute the target text layout into a preset database for similarity comparison, and if the comparison result indicates the existence of a nameplate image, record the detection image as a target image;

[0116] A final image acquisition module, configured to perform a perspective transformation matrix on the target image to obtain a final image.

[0117] Based on the nameplate area recognition and positioning system provided by an embodiment of the present invention, through the preprocessing and exposure adjustment of the initial image, the problems of overexposure or underexposure of the image caused by environmental light can be effectively reduced, thereby improving the recognition accuracy of nameplate information. Extracting the contours of the second image and cutting the set of regions to be detected helps to accurately locate the nameplate area, further improving the recognition precision and enhancing the accuracy and efficiency of nameplate recognition and positioning.

[0118] In one embodiment, the contour acquisition module includes:

[0119] An exposure image generation module, configured to process the first image through a tone mapping curve to obtain a set of exposure images;

[0120] A first parameter map acquisition module, configured to perform Laplacian filtering on each exposure image in the set of exposure images to obtain the filtering result value of each pixel in the exposure image, record the absolute value of the filtering result value as the first parameter value, and obtain a first parameter map according to the first parameter values of all pixels;

[0121] A second parameter map acquisition module, configured to calculate the mean value of each pixel in the R channel, G channel, and B channel of the pixel, obtain the standard deviation corresponding to the pixel according to the mean value, and obtain a second parameter map according to the standard deviations corresponding to all pixels;

[0122] A third parameter map acquisition module, which is used to calculate the brightness value corresponding to each pixel of the exposure image, obtain the exposure value corresponding to the pixel according to the brightness value, and obtain a third parameter map according to the exposure values corresponding to all pixels;

[0123] A second image generation module, which is used to obtain the weight value corresponding to the exposure image according to the first parameter map, the second parameter map and the third parameter map corresponding to the exposure image, and perform weighted summation on all exposure images according to the weight values corresponding to all exposure images to obtain a second image.

[0124] In one embodiment, the contour acquisition module further includes:

[0125] A predicted bounding box generation module, which is used to substitute the second image into a deep aggregation network to obtain a first multi-scale feature map, and perform center point, offset and direction prediction on the first multi-scale feature map to obtain a plurality of predicted bounding boxes;

[0126] An initial connection line generation module, which is used to perform corner point detection on each predicted bounding box to obtain a set of target corner points, and connect all points in the set of target corner points to obtain a set of initial connection lines;

[0127] A first connection line determination module, which is used to obtain a combination of initial connection lines with the largest formed area for the set of initial connection lines to obtain a set of first connection lines;

[0128] A region to be detected determination module, which is used to substitute the set of first connection lines and the second image into a connection line enhancement model to obtain a set of second connection lines, form a region to be detected according to the connection lines in the set of second connection lines, and obtain a set of regions to be detected according to the regions to be detected corresponding to each predicted bounding box.

[0129] In one embodiment, the region to be detected determination module includes:

[0130] An encoding optimization module, which is used to substitute the second image into a CNN backbone network to obtain a second multi-scale feature map, and perform encoding optimization on the second multi-scale feature map through a transformer encoder to obtain a target scale feature map;

[0131] A target query vector determination module, which is used to perform position encoding on the endpoint coordinates of each connection line in the set of first connection lines to obtain a first encoded value, and substitute the first encoded value into a multi-layer perceptron to obtain a target query vector;

[0132] A second connection line determination module, which is used to substitute the target query vector and the target scale feature map into a transformer decoder for decoding to obtain a target enhanced connection line, enhance the connection line according to the target enhanced connection line to obtain a target connection line, and obtain a set of second connection lines for all connection lines corresponding to the set of first connection lines.

[0133] In one embodiment, the target image acquisition module includes:

[0134] A target detection image determination module, configured to convert the detection image into a grayscale image, and convert the grayscale image into a binary image through an edge detection algorithm to obtain a target detection image;

[0135] A text region boundary determination module, configured to perform Hough transform detection on the target detection image to obtain the text region boundary in the target detection image, and determine the text recognition contour according to the text region boundary;

[0136] An image cutting module, configured to cut the target detection image according to the text recognition contour to obtain a target text image;

[0137] An image layout module, configured to obtain all target text images, perform text recognition to obtain a character label set, and layout all character labels in the character label set according to the recognition positions of all target text images to obtain a target text layout.

[0138] The above has described in detail an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the application of the present invention should still fall within the scope covered by the patent of the present invention.

Claims

1. A method for identifying and locating a nameplate area in a nameplate image, characterized in that: The method comprises: Acquire an initial image, and preprocess the initial image to obtain a first image; the initial image is an image including a nameplate area; Adjusting the exposure of the first image to obtain a second image, extracting contours in the second image, and obtaining a set of regions to be detected according to the extracted contours; For the set of regions to be detected, the second image is cut with the region to be detected to obtain a detection image, a text contour is extracted from the detection image to obtain a target text layout, the target text layout is substituted into a preset database for similarity comparison, and if the comparison result shows that there is a nameplate image, the detection image is recorded as a target image; Performing a perspective transformation matrix on the target image to obtain a final image; Adjusting the exposure of the first image to obtain the second image includes: Processing the first image by a tone mapping curve to obtain an exposure image set; For each exposure image in the exposure image set, Laplace filtering is performed on the exposure image to obtain a filtering result value of each pixel in the exposure image, an absolute value of the filtering result value is recorded as a first parameter value, and a first parameter map is obtained according to the first parameter values ​​of all pixels; For each pixel of the exposure image, the mean of the pixel on the R channel, the G channel and the B channel is calculated, the standard deviation corresponding to the pixel is obtained according to the mean, and the second parameter map is obtained according to the standard deviations corresponding to all pixels; For each pixel of the exposure image, calculate the brightness value corresponding to the pixel, obtain the exposure value corresponding to the pixel according to the brightness value, and obtain a third parameter map according to the exposure values ​​corresponding to all pixels; A weight value corresponding to the exposure image is obtained according to the first parameter map, the second parameter map and the third parameter map corresponding to the exposure image, and a weighted sum is performed on all exposure images according to the weight values ​​corresponding to all exposure images to obtain a second image.

2. The method for identifying and locating a nameplate area in a nameplate image according to claim 1, characterized in that: Extracting contours in the second image, and obtaining a set of regions to be detected according to the extracted contours comprises: Substituting the second image into a deep aggregation network to obtain a first multi-scale feature map, and performing center point, offset and direction prediction on the first multi-scale feature map to obtain multiple predicted bounding boxes; For each predicted bounding box, performing corner point detection on the predicted bounding box to obtain a target corner point set, and connecting all points in the target corner point set to obtain an initial connection line set; For the initial connection line set, obtaining the initial connection line combination that forms the largest area to obtain a first connection line set; The first connection line set and the second image are brought into the connection line enhancement model to obtain a second connection line set, a region to be detected is formed according to the connection lines in the second connection line set, and a set of regions to be detected is obtained according to the region to be detected corresponding to each predicted bounding box.

3. The method for identifying and locating a nameplate area in a nameplate image according to claim 2, characterized in that: Substituting the first connection line set and the second image into the connection line enhancement model to obtain the second connection line set comprises: Substituting the second image into a CNN backbone network to obtain a second multi-scale feature map, and performing encoding optimization on the second multi-scale feature map through a transformer encoder to obtain a target scale feature map; For each link in the first link set, position encoding is performed on the endpoint coordinates of the link to obtain a first encoding value, and the first encoding value is substituted into a multilayer perceptron to obtain a target query vector; The target query vector and the target scale feature map are brought into a transformer decoder for decoding to obtain a target enhanced link line, the link line is enhanced according to the target enhanced link line to obtain a target link line, and the target link lines corresponding to all the link lines in the first link line set are obtained to obtain a second link line set.

4. The method for identifying and locating a nameplate area in a nameplate image according to claim 1, characterized in that: Extracting the text contour of the detected image to obtain the target text layout includes: The detection image is converted into a grayscale image, and the grayscale image is converted into a binary image by an edge detection algorithm to obtain a target detection image; Performing Hough transform detection on the target detection image to obtain a text area boundary in the target detection image, and determining a text recognition contour according to the text area boundary; Performing text segmentation on the target detection image according to the text recognition contour to obtain a target text image; All target text images are acquired for text recognition to obtain a character label set, and all character labels in the character label set are laid out according to the recognition positions of all target text images to obtain a target text layout.

5. A nameplate area recognition and positioning system in a nameplate image, characterized in that: The system comprises: A first image acquisition module, used to acquire an initial image, and preprocess the initial image to obtain a first image; the initial image is an image including a nameplate area; A contour acquisition module, used to adjust the exposure of the first image to obtain a second image, extract contours in the second image, and obtain a set of areas to be detected according to the extracted contours; a target image acquisition module, for the set of regions to be detected, cutting the second image with the region to be detected to obtain a detection image, extracting text contours from the detection image to obtain a target text layout, substituting the target text layout into a preset database for similarity comparison, and if the comparison result shows that a nameplate image exists, recording the detection image as a target image; A final image acquisition module, used for performing a perspective transformation matrix on the target image to obtain a final image; The contour acquisition module comprises: An exposure image generation module, configured to process the first image using a tone mapping curve to obtain an exposure image set; a first parameter map acquisition module, configured to perform Laplace filtering on each exposure image in the exposure image set to obtain a filtering result value of each pixel in the exposure image, record an absolute value of the filtering result value as a first parameter value, and obtain a first parameter map according to the first parameter values ​​of all pixels; A second parameter map acquisition module is used to calculate the mean of each pixel of the exposure image on the R channel, the G channel and the B channel, obtain the standard deviation corresponding to the pixel according to the mean, and obtain the second parameter map according to the standard deviations corresponding to all pixels; A third parameter map acquisition module is used to calculate the brightness value corresponding to each pixel of the exposure image, obtain the exposure value corresponding to the pixel according to the brightness value, and obtain the third parameter map according to the exposure values ​​corresponding to all pixels; The second image generation module is used to obtain the weight value corresponding to the exposure image according to the first parameter map, the second parameter map and the third parameter map corresponding to the exposure image, and to obtain the second image by weighted summing up all the exposure images according to the weight values ​​corresponding to all the exposure images.

6. A nameplate area recognition and positioning system in a nameplate image according to claim 5, characterized in that: The contour acquisition module also includes: A prediction bounding box generation module, used for substituting the second image into a deep aggregation network to obtain a first multi-scale feature map, and performing center point, offset and direction prediction on the first multi-scale feature map to obtain multiple prediction bounding boxes; An initial connection line generation module is used to perform corner point detection on each predicted bounding box to obtain a target corner point set, and connect all points in the target corner point set to obtain an initial connection line set; A first connection line determination module is used to obtain, for the initial connection line set, a combination of initial connection lines that forms the largest area to obtain a first connection line set; The module for determining the area to be detected is used to bring the first connection line set and the second image into the connection line enhancement model to obtain a second connection line set, form the area to be detected according to the connection lines in the second connection line set, and obtain the area set to be detected according to the area to be detected corresponding to each predicted boundary box.

7. The system for identifying and locating a nameplate area in a nameplate image according to claim 6, characterized in that: The module for determining the area to be detected includes: A coding optimization module, used for substituting the second image into a CNN backbone network to obtain a second multi-scale feature map, and performing coding optimization on the second multi-scale feature map through a transformer encoder to obtain a target scale feature map; a target query vector determination module, configured to perform position encoding on the endpoint coordinates of each link in the first link set to obtain a first encoding value, and substitute the first encoding value into a multilayer perceptron to obtain a target query vector; The second connecting line determination module is used to bring the target query vector and the target scale feature map into the transformer decoder for decoding to obtain a target enhanced connecting line, enhance the connecting line according to the target enhanced connecting line to obtain a target connecting line, and obtain the target connecting lines corresponding to all the connecting lines in the first connecting line set to obtain a second connecting line set.

8. The system for identifying and locating a nameplate area in a nameplate image according to claim 5, characterized in that: The target image acquisition module comprises: The target detection image determination module is used to convert the detection image into a grayscale image, and convert the grayscale image into a binary image by an edge detection algorithm to obtain a target detection image; A text area boundary determination module is used to perform Hough transform detection on the target detection image to obtain the text area boundary in the target detection image, and determine the text recognition contour according to the text area boundary; An image cutting module, used for performing text cutting on the target detection image according to the text recognition contour to obtain a target text image; The image layout module is used to obtain all target text images for text recognition to obtain a character label set, and to layout all character labels in the character label set according to the recognition positions of all target text images to obtain a target text layout.

Images