| The logo recognition technology is in a stage of rapid development,and has been greatly improved in terms of image processing accuracy,reproducibility,flexibility,applicability,and information compression.However,in the actual application process,the development of this technology is limited by the actual operating environment,which is embodied in the large amount of calculation of the deep learning model and the slow processing speed of lightweight embedded devices,which is difficult to meet the real-time requirements of sign recognition;There are many and complicated noises in industrial production,which interfere with the texture feature information of signboards and affect the recognition accuracy of signboards.This subject takes the wall material production line as the research object.Aiming at the problems of low recognition rate of signboards in high-noise environments and slow running speed of recognition models on lightweight embedded devices,the research on embedded sign recognition technology in high-noise environments is launched.,The specific research content is as follows:In view of the impact of dust and noise on the identification of signboards in production,the existing defogging algorithms are analyzed,and the dark primary color a priori defogging algorithm suitable for removing haze-like dust is selected to reduce the influence of dust and noise on the identification of signboards.And use the two-dimensional gamma algorithm to correct the brightness of the image after dust removal,and solve the problem of dark brightness and unobvious sign characteristics after the image is dust removed.According to the method of direct part identification by DPM code,the identification plate is made by hollow engraving.The hollowing method can enhance the texture feature information of the signboard,and the concave-convex feature will not completely disappear with the color gradation of the signboard,and can provide characteristic elements for subsequent signboard recognition.In terms of sign edge extraction,the Canny edge detection algorithm is optimized to remove the internal Gaussian filter that has a large influence on the edge information,and the bilateral filter is used to supplement the noise reduction part to enhance the robustness of the edge information extraction.After edge extraction,the closed operation and open operation in mathematical morphology are used to expand and reduce the noise of the edge,remove the burrs in the edge,and screen out the signs that meet the requirements based on the length and width ratio of the sign.In order to further improve the recognition accuracy of the sign,considering the unique round hole feature information of the sign,on the basis of the classic LeNet-5,a combination of LeNet-5 and ellipse fitting is adopted.In model training,the model recognition result is compared with the ellipse fitting result,and the recognition result is considered accurate only when the number of ellipses in the model recognition result is consistent with the number of ellipses in the input image.While increasing a small amount of calculations,it enhances the classification ability of the LeNet-5 network model and improves the recognition accuracy of the sign.The experimental analysis shows that the combination of the dark primary dehazing algorithm and the gamma brightness correction algorithm removes the interference of haze dust on the sign;the optimized Canny edge detection algorithm is combined with mathematical morphology to achieve high Accurate extraction of signs in noisy environments;the LeNet-5model optimized by ellipse fitting has the advantages of less computation,fast speed,and high recognition rate,and it can run on the embedded device Raspberry Pi.The embedded sign recognition technology studied in this subject realizes the recognition of signboards in a highnoise environment,and has the advantages of fast running speed and high reliability. |
