Scene Text Location Algorithm Research Based On Maximally Stable Extremal Regions

There is lots of semantic information in natural scene image,and it is an important complement to the scene and a key clue to understanding the content of scene.The popularization of smartphone,tablet PC and digital camera brings the convenience of capturing high quality images.Text detection in natural scene image can not only help people to have a deep understanding of the scene but also play an important role in the information retrieval system and visual aid system.Accurate text region location is the prerequisite for text information extraction.As for the complex background and lighting conditions,as well as multi-scale,multi-lingual and multi-directional,text detection in natural scene images is a challenging task.There has been a lot of scene text location algorithms currently.Based on exploring the existing algorithms,a new algorithm framework based on MSER is proposed.The main contributions of this thesis are as follows.1.Take the problem of repeated detection of MSER into account,an algorithm based on non-maximum suppression of variations is designed.Firstly,extract MSERs as the candidate regions from the positive and negative channels.Then a rule based on the variation of MSER is designed to delete the repeated nesting regions.2.The edge-enhanced MSER is designed to solve the problem of edge adhesion between the adjacent characters in low resolution images.Based on extracting the edge-enhanced MSER as the character candidates,some heuristic rules are designed to delete the non-character regions initially.Then the application of stroke width transformation and SVM classifier are used to delete the rest non-character regions.3.In order to express the complete semantic information,the character regions are grouped into text line based on a new clustering rule.Finally,the proposed method is evaluated on three public datasets,the ICDAR 2003 Dataset,the ICDAR 2013 Dataset and SVT dataset.And the experiment results show that the proposed method based on MSER and S VM can achieve good performance.