Description

We propose a hybrid text detection and localization approach by combining region-based and CC-based information. At the pre-processing stage, a text region detector is designed to detect text regions in each layer of the image pyramid and project the text confidence and scale information back to the original image, scale-adaptive Binarization is then applied to generate candidate text components. At the connected component analysis (CCA) stage, a CRF model combining unary component properties (including the text confidence) and binary contextual component relationships is used to filter out non-text components. At the last stage, neighboring text components are linked with a learning-based classification – regression tree. To design a text region detector and group text components into text lines/words will be effectively analyzed.