Detection And Reading Out Of Scene Text Images

摘要

Scene text detection method based on connected components clustering and nontext filtering is proposed in this paper. Text region can be extracted and readout using this system and can be used in a number of applications to support visually impaired people. Two machine learning classifiers are used (one to generate candidate regions, and the second one to extract non text ones). Connected components, extracted using maximally stable extremal region algorithm are partitioned in to clusters. Candidate regions are generated from these clusters. For clustering AdaBoost classifier is used and it clusters CC by determining adjacency relationship using their pair wise relation. In the second stage candidate region will be normalized. And from these geometrically normalized images, binary images will be generated. In this paper Kernel Discriminate analysis method is proposed to filter the text blocks. These text regions will be highlighted, and will be readout the text.