Research On Text Detection For Nature Scene Images

With the development of society,more and more useful information is transmitted in images and videos,so it is very important to extract these information.Texts in images contain much special and important information,therefore are widely studied and applied to be detected and extracted.Although the traditional OCR technology becomes mature,the technology of text detection and recognition in natural images can not meet the need of practical application.The thesis mainly studies the text detection in natural images included the following aspects:1.The thesis briefly introduces the domestic and foreign research of the technology of text detection in natural images based on the traditional manual design features and depth learning.The open datasets are introduced briefly.The Chinese text and Synthetic-Data are made.Most of the algorithms in the thesis are trained and tested in the ICDAR2013 and the Chinese text.2.Firstly,The candidate regions are extracted by the combination of SWT and MSER.Secondly,the final results are obtained by mutual authentication and some rules.The method is improved for features of Chinese text.3.The method named TextBoxes is improved based on DSSD,and the test results are better than before.The feature of corner is applied in the text detection of the RRPN algorithm to solve the problem of arbitrary-oriented scene text detection.4.With the development of an intelligent electric system,greater demands for the recognition of nameplates is put forward.In the paper,the scene text detection is applied in the detection of nameplates,and good test result is obtained.