Research On Chinese Text Detection Algorithm For Natural Scenes Based On YOLO

With the rapid development of Internet technology,text detection has important applications in the fields of intelligent transportation and video retrieval.This thesis summarizes the related research and technical foundation of the existing text detection technology.Aiming at the problems of text inclination,illumination,complex background structure in scene detection,and the difficulty of Chinese detection,combining the advantages of deep learning and traditional text detection methods,two Chinese text detection algorithms are proposed to mine the deep information of natural scene picture text semantics.The main work contents are as follows.(1)Aiming at the problems of low accuracy,slow speed,and detection when text lines are arranged in any direction in traditional natural scene text detection methods,it is proposed to add MSER on the basis of YOLO,and introduce GIo U to improve the border regression loss function.The improved model can make full use of the rapidity of the YOLO model and the advantages of the MSER algorithm in the detection of tilt angle,to achieve fast and accurate detection of text lines in any direction.Experiments show that the F-score of the algorithm on the public data set reaches 53.1.(2)In order to detect difficulties caused by high noise such as background blur and low contrast in natural scenes,directly improve the YOLO deep convolutional neural network model.The main idea of the algorithm: add a rotation proposal module(RPM)to the YOLO deep convolutional neural network model,redesign the loss function for angles,adjust and optimize the YOLO feature pyramid network structure and initialize the anchor box parameters,so that the YOLO model comes with rotation angle detection function.Experiments show that compared with(1)algorithm,the accuracy of this algorithm is improved by 2.8%,and the F-score reaches 54.6.