Surface Character Recognition Method for Composite Materials

Abstract

The surface character recognition method of the composite material of the present invention relates to the technical field of character recognition. The image is processed by using an edge-preserving filtering algorithm, and then gray-scale stretching is performed by using a penalty factor to obtain a detail weight matrix. The image is threshold-truncated to obtain a high-reflectance region weight matrix. The image is processed by using a top-hat transform to obtain image A, and then processed by using an edge-preserving filtering algorithm A to obtain low-frequency information. Image F and image A are subjected to joint bilateral filtering processing to obtain image Anr. The low-frequency information and image Anr are fused by using the detail weight matrix and the high-reflectance region weight matrix to obtain image Afinal. After Gaussian filtering is performed on image Afinal, the character region is screened out through horizontal projection and vertical projection, and the character information is obtained through a CRNN deep learning neural network, which solves the problem of low accuracy of surface character recognition for composite materials in the prior art. The present invention is applicable to surface character recognition of composite materials.

Description

Technical Field

[0001] The present invention relates to the technical field of character recognition, and particularly to a method for recognizing surface characters of composite materials. Background Art

[0002] With the continuous development of machine vision and artificial intelligence technologies, the automatic optical inspection (AOI) technology has been widely applied to the inspection of industrial finished products, such as defect detection, optical character recognition, target detection, etc. In the surface inspection in the field of machine vision, most of the optical character recognition, bar code detection, etc. are for the surface inspection of single light-transmitting materials or single light-reflecting materials, whose optical properties are relatively single, and at the same time, the image background is relatively pure. With the further development of market demands, packaging materials are becoming increasingly complex, and delicate packaging materials such as PET, electroplated materials, and visual anti-counterfeiting stickers are emerging continuously, making the surface inspection of non-single-attribute materials of increasing importance in practical applications. The demand for surface inspection of objects containing mirror-reflecting material surfaces, non-mirror-reflecting material surfaces, and light-transmitting and transparent material surfaces is increasing day by day. However, due to the great difficulty in realizing it in terms of optical imaging and image processing algorithms, the research in this aspect has received extensive attention.

[0003] For example, in the packaging industry, in order to improve the product appearance and anti-counterfeiting effect, production enterprises often use lasers to print specific characters in the form of multiple lines of text onto the overlapping areas of various composite materials. In the production inspection stage, these characters must be rechecked to ensure that no defective products flow into the market. The traditional manual method completely fails to meet the rapid comparison under high-speed production. However, the multi-material composite materials used in the packaging industry will cause complex situations such as strong reflection points, black-and-white alternating areas, and digital stitching in the images collected by the camera, making it difficult for traditional character recognition technologies to accurately recognize characters from these complex background interferences.

[0004] For the imaging of surfaces of composite materials including reflection, light transmission, etc., the current methods mainly rely on multiple imaging, that is, adopting targeted imaging schemes for materials with different optical properties, and then combining and piecing together all the imaging results through algorithms to form an overall image. This method of multiple imaging results in a complex hardware structure, multiple imaging stations, and high product costs. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for recognizing surface characters of composite materials, so as to solve the problem of low accuracy in recognizing surface characters of composite materials in the prior art.

[0006] The technical solution adopted by the present invention to solve the above technical problem: The method for recognizing surface characters of composite materials includes the following steps:

[0007] S01. Obtain the surface image F of the composite material;

[0008] S02. Process the image F using an edge-preserving filtering algorithm to obtain the Fbase image, and perform gray-scale stretching on Fbase in combination with a penalty factor to obtain the detail weight matrix Fdetail; perform threshold truncation on the image F to obtain the high-specular-reflection region weight matrix M; process the initial image F using a top-hat transform to obtain the image A, and process the image A using the edge-preserving filtering algorithm A to obtain the low-frequency information Abase; perform joint bilateral filtering on the image F and the image A to obtain the image Anr;

[0009] S03. Fuse the low-frequency information Abase and the image Anr using the detail weight matrix Fdetail and the high-specular-reflection region weight matrix M to obtain the image Afinal;

[0010] S04. Perform Gaussian filtering on the image Afinal and then screen out the character region through horizontal projection and vertical projection;

[0011] S05. Obtain the character information of the screened character region through a CRNN deep learning neural network.

[0012] Further, in step S01, use a character imaging system to obtain the surface image F of the composite material. The character imaging system includes an image acquisition device, a side illumination device, and a top illumination device. The image acquisition device obtains the diffuse reflection light of the specular reflection region and the dark surface region on the surface of the composite material through the side illumination device, and obtains the diffuse reflection light of the transparent region on the surface of the composite material through the top illumination device.

[0013] Further, in step S02, the detail weight matrix where ε is the penalty factor.

[0014] Further, in step S02, the method for obtaining the high-specular-reflection region weight matrix M is as follows: If the gray value of a pixel point in the image F is greater than the threshold, the weight of the pixel point is 1; otherwise, the weight of the pixel point is 0. Arrange and combine the weights of all pixel points in the image according to the pixel point coordinates to form a matrix, and this matrix is the high-specular-reflection region weight matrix M.

[0015] Further, in step S03, the fusion formula is: Afinal = (E - M) × Anr × Fdetail + M × Abase, where E represents the identity matrix.

[0016] Further, in step S04, the process of screening the character region is as follows: Set the gray value range and error threshold of the image Afinal, count the number of pixel gray values corresponding to the pixel coordinates in the horizontal direction of the image Afinal within the gray value range, and calculate the difference in the number of pixel gray values corresponding to adjacent pixel coordinates within the gray value range in ascending order. When the difference is greater than the error threshold for the first time, the minimum value in the corresponding adjacent pixel coordinates is the starting pixel coordinate in the horizontal direction. When the difference is greater than the error threshold for the last time, the maximum value in the corresponding adjacent pixel coordinates is the ending pixel coordinate in the horizontal direction. Count the number of pixel gray values corresponding to the pixel coordinates in the vertical direction of the image Afinal within the gray value range, and calculate the difference in the number of pixel gray values corresponding to adjacent pixel coordinates within the gray value range in ascending order. When the difference is greater than the error threshold for the first time, the minimum value in the corresponding adjacent pixel coordinates is the starting pixel coordinate in the vertical direction. When the difference is greater than the error threshold for the last time, the maximum value in the corresponding adjacent pixel coordinates is the ending pixel coordinate in the vertical direction. The rectangle formed by the starting pixel coordinate and ending pixel coordinate in the horizontal direction and the starting pixel coordinate and ending pixel coordinate in the vertical direction is the screened character region.

[0017] Advantages of the present invention: The present invention makes full use of the different illumination characteristics of reflective and light-transmitting materials. Through the side illumination device and the top illumination device, the image acquisition device fully obtains the surface information of the light-transmitting material and the reflective material, that is, image F. The image F is processed using the edge-preserving filtering algorithm to obtain the Fbase image, and the Fbase is subjected to gray-scale stretching in combination with the penalty factor to obtain the detail weight matrix Fdetail. The image F is thresholded to obtain the high-reflectance region weight matrix M. The initial image F is processed using the top-hat transform to obtain image A, and image A is processed using the edge-preserving filtering algorithm A to obtain the low-frequency information Abase. The image F and image A are subjected to joint bilateral filtering to obtain image Anr. The low-frequency information Abase and image Anr are fused using the detail weight matrix Fdetail and the high-reflectance region weight matrix M to obtain image Afinal. After Gaussian filtering of image Afinal, the character region is screened through horizontal projection and vertical projection, and the character information is obtained through the CRNN deep learning neural network, solving the problem of low accuracy in surface character recognition of composite materials in the prior art. Description of the Drawings

[0018] Attached Figure 1 is a schematic flow chart of the method for recognizing surface characters of the composite material of the present invention. Detailed Embodiments

[0019] The method for recognizing surface characters of the composite material of the present invention, as attachedFigure 1 As shown in the figure, it includes the following steps:

[0020] S01. Obtain the surface image F of the composite material;

[0021] Specifically, use a character imaging system to obtain the surface image F of the composite material. The character imaging system includes an image acquisition device, a side lighting device, and a top lighting device. The image acquisition device obtains the diffuse reflection light of the specular reflection area and the dark surface area on the surface of the composite material through the side lighting device, and obtains the diffuse reflection light of the transparent area on the surface of the composite material through the top lighting device. Thus, the image acquisition device fully obtains the surface information of the light-transmitting material and the light-reflecting material. When obtaining the surface image F of the composite material, the image can be made clearer by adding a resolution plate.

[0022] S02. Process the image F using an edge-preserving filtering algorithm to obtain the Fbase image, and perform gray-scale stretching on Fbase in combination with a penalty factor to obtain the detail weight matrix Fdetail; perform threshold truncation on the image F to obtain the high-reflection area weight matrix M; process the initial image F using a top-hat transform to obtain the image A, and process the image A using the edge-preserving filtering algorithm A to obtain the low-frequency information Abase; perform joint bilateral filtering on the image F and the image A to obtain the image Anr;

[0023] Specifically, the detail weight matrix where ε is the penalty factor; the high-reflection area weight matrix M is obtained as follows: if the gray value of a pixel point in the image F is greater than the threshold, the weight of the pixel point is 1, otherwise the weight of the pixel point is 0. Arrange and combine the weights of all pixel points in the image according to the pixel point coordinates to form a matrix, and the matrix is the high-reflection area weight matrix M.

[0024] S03. Use the detail weight matrix Fdetail and the high-reflection area weight matrix M to fuse the low-frequency information Abase and the image Anr to obtain the image Afinal;

[0025] Specifically, the fusion formula is: Afinal = (E - M) × Anr × Fdetail + M × Abase, where E represents the identity matrix. Since the low-frequency information Abase is used in the fusion formula, the basic information of the image will not be lost. Moreover, in the fusion process, the high-reflection area weight matrix of the original image is obtained by threshold truncation to eliminate the interference of high reflection of the image and reduce the interference of small branches on the text area positioning.

[0026] S04. Perform Gaussian filtering on the image Afinal and then screen out the character area through horizontal projection and vertical projection;

[0027] Specifically, the process of screening the character region is as follows: Set the grayscale value range and error threshold of the image Afinal. The grayscale value range is preferably 200 to 256 to improve the robustness of the screened character region. Count the number of pixel grayscale values corresponding to the pixel coordinates in the horizontal direction of the image Afinal within the grayscale value range, and calculate the difference in the number of pixel grayscale values corresponding to adjacent pixel coordinates within the grayscale value range in ascending order. When the difference is greater than the error threshold for the first time, the minimum value among the corresponding adjacent pixel coordinates is the starting pixel coordinate in the horizontal direction. When the difference is greater than the error threshold for the last time, the maximum value among the corresponding adjacent pixel coordinates is the ending pixel coordinate in the horizontal direction. Count the number of pixel grayscale values corresponding to the pixel coordinates in the vertical direction of the image Afinal within the grayscale value range, and calculate the difference in the number of pixel grayscale values corresponding to adjacent pixel coordinates within the grayscale value range in ascending order. When the difference is greater than the error threshold for the first time, the minimum value among the corresponding adjacent pixel coordinates is the starting pixel coordinate in the vertical direction. When the difference is greater than the error threshold for the last time, the maximum value among the corresponding adjacent pixel coordinates is the ending pixel coordinate in the vertical direction. The rectangle formed by the starting pixel coordinate and ending pixel coordinate in the horizontal direction and the starting pixel coordinate and ending pixel coordinate in the vertical direction is the screened character region.

[0028] S05. Obtain character information from the screened character region through a CRNN deep learning neural network.

Claims

1. Method for surface character recognition of composite material, characterized in that, It includes the following steps: S01. Use a character imaging system to obtain the surface image F of the composite material. The character imaging system includes an image acquisition device, a side illumination device, and a top illumination device. The image acquisition device obtains the diffuse reflection light of the specular reflection area and the dark surface area on the surface of the composite material through the side illumination device, and obtains the diffuse reflection light of the transparent area on the surface of the composite material through the top illumination device; S02. Use an edge-preserving filtering algorithm to process the image F to obtain the Fbase image, and perform gray-scale stretching on the Fbase image in combination with a penalty factor to obtain the detail weight matrix Fdetail; perform threshold truncation on the image F to obtain the high-reflectance area weight matrix M; use a top-hat transform to process the image F to obtain the image A, use an edge-preserving filtering algorithm to process the image A to obtain the low-frequency information Abase; perform joint bilateral filtering on the image F and the image A to obtain the image Anr; S03. Use the detail weight matrix Fdetail and the high-reflectance area weight matrix M to fuse the low-frequency information Abase and the image Anr to obtain the image Afinal; S04. Perform Gaussian filtering on the image Afinal and then screen out the character area through horizontal projection and vertical projection; S05. Obtain the character information by passing the screened character area through a CRNN deep learning neural network.

2. The method for surface character recognition of the composite material according to claim 1, wherein In step S02, the detail weight matrix where ε is a penalty factor.

3. The surface character recognition method of the composite material according to claim 1, wherein In step S02, the high-reflectance area weight matrix M is obtained as follows: If the gray value of the pixel point in the image F is greater than the threshold, the weight of the pixel point is 1, otherwise the weight of the pixel point is 0. Arrange and combine the weights of all pixel points in the image according to the pixel point coordinates to form a matrix, and this matrix is the high-reflectance area weight matrix M.

4. The method for surface character recognition of the composite material according to claim 1, wherein, In step S03, the fusion formula is: Afinal = (E - M) × Anr × Fdetail + M × Abase, where E represents the identity matrix.

5. The surface character recognition method of the composite material according to claim 1, wherein, In step S04, the process of screening out the character region is as follows: Set the grayscale value range and error threshold of the image Afinal, count the number of pixel grayscale values corresponding to the pixel coordinates in the horizontal direction of the image Afinal within the grayscale value range, and calculate the difference in the number of pixel grayscale values corresponding to adjacent pixel coordinates within the grayscale value range in ascending order. When the difference is greater than the error threshold for the first time, the minimum value among the corresponding adjacent pixel coordinates is the starting pixel coordinate in the horizontal direction. When the difference is greater than the error threshold for the last time, the maximum value among the corresponding adjacent pixel coordinates is the ending pixel coordinate in the horizontal direction; Count the number of pixel grayscale values corresponding to the pixel coordinates in the vertical direction of the image Afinal within the grayscale value range, and calculate the difference in the number of pixel grayscale values corresponding to adjacent pixel coordinates within the grayscale value range in ascending order. When the difference is greater than the error threshold for the first time, the minimum value among the corresponding adjacent pixel coordinates is the starting pixel coordinate in the vertical direction. When the difference is greater than the error threshold for the last time, the maximum value among the corresponding adjacent pixel coordinates is the ending pixel coordinate in the vertical direction; The rectangle formed by the starting pixel coordinate and ending pixel coordinate in the horizontal direction and the starting pixel coordinate and ending pixel coordinate in the vertical direction is the screened-out character region.

Images