Tyre Imprint Character Recognition Based On Deep Learning

The mark on the tire tread should contain information such as origin information,the manufacturer’s code and production date according to the Department of Transportation standard.These identification characters are directly imprinted on the tire through the mold,known as imprint characters.Different from the traditional optical characters,the contrast between the imprinted characters and the background region is low,which will cause difficulty in character recognition,and the efficiency of manual recognition is low,and easy to misjudge.With the improvement of automation production efficiency and the development of science and technology,automatic integrated identification of embossed characters has become in urgent demand.In this paper,the efficient and accurate identification of tire DOT characters is studied.Based on the basic theoretical knowledge and target detection model of convolutional neural network,the function of DOT identification detection and character recognition is realized.This paper mainly includes four parts:tire image preprocessing,DOT location,tire DOT segmentation and tire identification character recognition.In tire image pretreatment,in order to minimize the interference of background on the character,extract the interested in tyre round ring area,we attempted to find the inner circle based on the gradient Hough transform,select the partition type random pixels,and calculate the distance of each point to the center of the circle,which will be ordered a cumulative value as the most cylindrical radius.Compared with the concentric circles detected by the projection method,the detection time of this method is greatly shortened and the efficiency is improved by 23.9%.The extracted region is reconstructed based on the rectangular distribution.Meanwhile,the complexity of subsequent character segmentation is correspondently reduced.In the DOT identification location detection,based on the Faster R-CNN algorithm,the Bilinear interpolation method for subsampling,so as to improve the detection accuracy of the edge information,and filtering by threshold method.The accuracy of the nearest neighbor interpolation method,bilinear interpolation method and the Faster R-CNN algorithm of the threshold method used in this paper were compared with 79.5%,76.2% and 93.7% respectively.Therefore,the threshold method Faster R-CNN algorithm in this paper can greatly improve the accuracy of DOT detection.In the DOT segmentation and tire identification character recognition.First,the tire imprint character segmentation algorithm of projection method was adopted,then the 36 characters were extracted based on VGG19 network structure,and achieve the character recognition.The final character recognition accuracy reached 98.16%.This method overcomes the defect of unstable accuracy of imprint character recognition and achieves the experimental result of 95% accuracy.