Research On Chinese Text Localization Methods In Natural Scene Images

Text in natural scene images contains a wealth of information, thus extracting this informationplays a great role in describing and understanding natural scenes. Recently, text extraction of naturalimages has effectively improved the development of fields such as content-based image and videoretrieval, network security, visual and tourist auxiliary systems. Currently, text localization methodsare still far from reaching actual needs of users, and most of them are aimed at locating English text.And features of Chinese and English characters have great difference. Thus this paper mainly doesresearch on Chinese text location and eventually proposes two solutions to locate Chinese texteffectively. These two methods both contain four parts: pre-processing, candidate text regiongeneration, feature extraction and text region classification.(1) Pre-processing and candidate text region generation. In the part of pre-processing, first itextracts edge information and then binaries it with improved Niblack algorithm to get better edgeinformation. In the part of candidate text regions generation, first it eliminates long lines andisolated noise. Then, it takes morphological processing connected component analysis to filter outnon-text regions. Experimental results show that candidate text region generation could eliminatemost of non-text regions effectively and obtain the candidate text regions quickly and completely.(2) Feature extraction. This paper presents PHOG-Gabor4feature to classify Chinese characterwhere multi-scale PHOG features is used to describe contour information and its spatial distribution,multi-scale and direction Gabor features to describe structure, and four other kinds of features tosupplement the description of Chinese characters. Take the joint features as input to train classifiers,and experimental results show that the extracted features can effectively describe texture features ofChinese characters.(3) Text region classification. In the first location method, it uses SVM and Boosting tree asclassifier respectively. In the second location method, it proposes using Labeled-LDA as classifierto improve the classification result of discriminative models. Experimental results show thatclassification accuracy of Boosting tree is slightly better than that of SVM while classificationaccuracy of Labeled-LDA is much better than these two models.Finally, based on the research and analysis of ICDAR2003contest image library for Englishtext location, this paper itself establishes an image library for Chinese text location at the same level. This paper tests the proposed two Chinese text location methods and other existing locationmethods on this library. The location method based on discriminative model has a precision rate of0.83and recall rate of0.86. And location method based on generation model Labeled-LDA has aprecision rate of0.87and recall rate of0.90. Experimental results show that the two locationmethod are better than that of other referred papers, especially the method based on Labeled-LDAcan further improve the location accuracy rate with great robustness.