Scene Text Detection With Adaptive Line Clustering

Text in scene images usually conveys valuable information and text localization plays an important role in many practical applications,hence it has received signifi-cant attention in the last decades.The prevalent scene text detection approach follows four sequential steps comprising character candidate detection,false character candi-date removal,text line extraction,and text line verification.However,errors occur and accumulate throughout each of these sequential steps which often lead to low detection performance.We propose a scene text detection system which can maintain a high recall while achieving a fair precision.In our method,no character candidate is eliminated based on character-level features.Instead,we just assign a probability to each one.After extracting line candidates,the system assigns three line-level probability values to each line.The final decisions are made according to the line candidate clustering of the current image.In the meanwhile,we propose new methods for character candidate detection and text line extraction,in order to detect text lines with arbitrary orientations in both English and Chinese.To be specific,we take advantage of the properties of Chinese characters and design a hybrid method which applies a sliding window method after the connected component analysis.This strategy has better performances for the Chinese characters and barely affects the extraction of the English characters.In addition,we make full use of the consistency in direction and adopt a graph partitioning method to get lines with arbitrary orientations.The proposed system has been evaluated on two text detection benchmarks:ICDAR-13 and MSRA-TD500.Compared with other published methods,it achieves the state-of-the-art performance.