Research On Text Detection And Curved-Text Rectification Of Chinese In Scene Images

Chinese character,as one of the oldest character and most widely used languages in the world,is the soul bond of the Chinese nation and the cultural treasure of the Chinese nation for thousands of years,and plays a decisive role in all aspects of human production,life and development.Chinese character is commonly found around modern cities,such as road signs,billboards,license plates,posters and so on.In view of the large number and wide serviceable range of Chinese characters,it is of great significance to realize the automatic detection and recognition of Chinese characters in natural scene images.However,on the one hand,natural scene images are affected by complex background,noise,occlusion and other unfavorable factors,which brings great difficulties and challenges to the detection and recognition of Chinese.On the other hand,because of the existence of irregular text,such as text suffering perspective deformation and text arranged along curves,it is more difficult to detect and recognize the text in natural scene images.Research on text detection and rectification of irregular text could promote the automatic detection and recognition of Chinese text in scene images.This thesis will focus on the text detection of Chinese in scene images and the rectification of text arranged along curves.The main work includes:(1)A text detection method based on Maximally Stable Extremal Regions(MSER)and Stroke Width Transform(SWT)was used.The text candidate regions were detected by MSER,the non-text candidate regions were removed according to the geometric characteristics of the text and SWT.The bounding box of each text region was determined.The boxes were merged according to the overlap rate of the adjacent ones to detect the whole text region.(2)For Chinese text suffering perspective deformation and Chinese text arranged along curves,a segmentation method based on connected regions was proposed.The center and radius of the bounding circle of each region were used to describe the position and size of each part of the character.Rules were established to binding different connected regions belonging to one Chinese character.The experiments show that the Chinese characters suffering perspective deformation and the Chinese characters arranged along curves were well segmented using the proposed method.(3)A method was proposed to straighten the curved Chinese text taking advantage of the characteristics of the horizontal and the vertical strokes.The tilt angle of the horizontal or the vertical strokes was first determined,and then the tilt angle of each character was determined considering the direction of the text line.For characters without horizontal and vertical strokes,the tilt angle could be determined by its adjacent characters.Experiments show that curved Chinese texts were well straightened using the proposed method.