A Study On Deep Learning Based Chinese Text Detection And Recognition

Text detection and recognition applications can be seen everywhere in our daily life,such as ID card recognition,driver's license recognition,etc.,which greatly facilitates people's lives.In different scenarios,texts are often of various forms,scales,and semantic backgrounds.It is necessary to design detection and recognition algorithms accordingly to achieve better performance.This thesis is focusing on the title detection and recognition of the Republic of China newspapers,it is conducive to transforming these newspapers into digital ones that can be searched and queried,furthermore,mining their historical value.After a thorough study of deep learning based text detection and recognition works,we propose to improve existing algorithms based on the characteristics of the Republic of China newspapers,and design the following algorithm scheme.The quality of raw images of the Republic of China newspapers is affected by the shooting environment and the professionalism of the photographer,however,it still con-tains black and white edges,and may have a certain degree of tilt.This not only increases unnecessary storage consumption,but also affects the e-reading experience.We propose a layout analysis method with line segmentation detection algorithm and projection method as the core.By writing a series of heuristic rules,preprocessing operations including black and white edge removal and image deskew are designed.In the title detection stage,this thesis adopts and improves the single-stage anchor-free detector EAST.EAST has fast inference speed,and its feature of anchor free elimi-nates the tedious work of re-adjusting the scales and ratios of the anchors when applied to a new task,which together are conducive to the model's deployment,improvements,and adaptation.The printed traditional characters used in the Republic of China newspapers have complex structures and varied distortion,therefore,we improve the feature extractor to learn more distinguishing image features.In addition,the design of the loss function has a great impact on the performance of the model,we propose to change the loss func-tion of EAST so it is able to achieve better detection performance in the Republic of China newspaper scenario.CRNN model is used in the title recognition module to make full use of the visual and semantic information in the title text sequence.The Republic of China newspaper recognition dataset has limited training samples,we generate a great number of fake data which is close to the real scene to alleviate this phenomenon.Considering that the se-mantic directions of vertical and horizontal titles in the Republic of China newspapers are different,and CRNN has a fixed semantic direction,we propose to apply two models to handle this situtaion.We verify the effectiveness of the above methods and improvements through a lot of experiments.Our work is a good exemplar to Chinese text detection and recognition tasks in other scenarios.