A Robust LGP-based Approach to Detect Text in Complex Scene

• 作 者: SiqiWANG;TianchenLI;EnZHU;JianpingYIN;
• 机 构: School of Computer Science,National University of Defense Technology,Changsha 410073,China
• 出 处: 2014全国理论计算机科学学术年会
• 发表日期: 2004年期

　　This paper presents a new coarse-to-fine scene text detection approach which can detect scene text fast and robustly.In the coarse stage,the Maximally Stable Extremal Regions (MSER) and geometric filtering are used to select candidate regions.Subsequently,Adaboost-based boosting classifiers are utilized for the further removal of non-text candidates.In the fine stage,we introduce novel vsLGP and bpLGP features,which turn out to be more insensitive to background contrast reversal or local intensity variations than other traditional features,to represent the text texture in terms of both overall profile and details.Then two weighted kernel based Extreme Learning Machine (ELM) using vsLGP and bpLGP are trained as the final classifiers at an extremely fast speed.Then they are cascaded for the sake of better detection performance.Eventually,detected text boxes at different scales of image are merged to obtain the final bounding box of text regions.Experiment shows the proposed approach yields competitive results on ICDAR 2003 Dataset in comparison with other state-of-the-art methods.