Research On Optical Character Recognition Based On Recurrent Neural Network

With the development of modern computer technology,application scenarios such as car plate recognition,electronic document recognition,business receipt management,and photo translation can be seen everywhere.If we can use optical character recognition(OCR)technology to recognize text from images automaticlly,people's work efficiency will be greatly improved,and quality of life will be better.Therefore,OCR technology has become one of the important development directions in computer vision technology.In natural scenarios,text images may have different sizes,fonts,colors,arrangement methods,etc.due to inaccurate focus,insufficient light,too much noise,angle tilt,and image distortion,which would cause poor text recognition accuracy.Therefore,to solve the problem of poor image quality caused by text tilt in natural images,this paper proposes an integrated optical detection and recognition system,and proposes a method of text image detection and recognition based on deep learning neural network,and use recurrent neural network to help recognizing text images.The specific contents are as follows: Firstly,the text detection and correction method based on YOLO-text network is proposed for the text detection module.The object detection network YOLOv3 is selected as the basis,and the shortcomings of YOLOv3 which not suitable for text detection are optimized and improved.Based on YOLO-text,in order to solve the problem of non-horizontal text areas in natural scene images,a method of detecting the letter bounding box using "calculus thinking",splicing all the letters into a complete string is proposed.At the same time,the "angle regression thinking" is proposed to identify the entire string bounding box and then correct the angle of the string in the image.Secondly,a text recognition method based on STN-text network is proposed for the text recognition module.The network combines CRNN and Spatial Transformer Network.The network corrects the input images of irregularities,deformations and distortions and then recognizes them,to improve the recognition accuracy.In addition,this paper proposes a method of adaptively filling black areas on both sides of the image,instead of direct stretching them in the image preprocessing procedure,which reduces the recognition inaccuracy caused by excessive deformation during the uniform scaling.This paper combines the text detection module and the text recognition module,proposing a complete OCR system.In the experiment comparison,the YOLO-text-based text detection module has improved the F-Measure by 3% compared with other algorithms;the STN-textbased text recognition module does not exceed CRNN.But compared with the traditional text recognition method,it shows obvious advantages;the complete OCR system combined with the detection and recognition module has a certain improvement in F-Measure compared with other algorithms,and has better performance.