Research On Scene Text Detection And Recognition Based On Deep Learning

Text in natural scene images contains important high-level semantic information,which is helpful to analyze and understand the corresponding environment.In recent years,Scene Text Recognition(STR)has become an active research topic in computer vision.Although the traditional Optical Character Recognition(OCR)method has a high recognition rate for characters in fixed documents,it cannot be applied to the actual scene image text recognition.The rapid development of Deep Learning(DL)provides more research schemes for scene text detection and recognition.However,due to the characteristics of different text shapes,changeable angles and complex background information,the research on text detection and recognition in natural scenes still faces enormous challenges.Based on deep learning technology,this paper proposes a novel algorithm for text detection and recognition in natural scenes.The specific work are as follow:1.To solve the problem of complex scene text detection,an end-to-end text detection method STDP(Scene Text Detection Perception)is designed,and the text detection method based on semantic segmentation is used to learn the reading order and boundary information of potential text.The regression strategy is adopted to adjust the wrong detection region to solve the problem of regional deviation in text detection.A new Text Shape Transformation Module(TSTM)is designed for irregular texts with feature regions,which can transform the detected feature regions into regular shapes without additional parameters,and can better solve the incompatibility between text detection and recognition,and improve the recognition efficiency and accuracy.2.To solve the problem of natural scene text recognition,a new scene text recognition method is designed on the basis of Convolutional Recurrent Neural Network(CRNN),which combines attention mechanism with connectionist temporal classification,(CTC)can solve the problems of text information loss and slow training speed caused by single decoding,and realize the rapid convergence of the network through joint decoding training.3.In this paper,the TSTM module is used to connect the detection method with the recognition method,forming an end-to-end detection and recognition network,which is verified and analyzed on different data sets.Experiments show that the accuracy of the text recognition algorithm ACTR proposed in this paper reaches98.89%,which greatly surpasses the single decoding recognition algorithm.Compared with the latest end-to-end detection algorithm,the proposed end-to-end text detection method has a significant improvement in detection index.