Aggregating local context for accurate scene text detection

Abstract

Scene text reading continues to be of interest for many reasons including applications for the visually impaired and automatic image indexing systems. Here we propose a novel end-to-end scene text detection algorithm. First, for identifying text regions we design a novel Convolutional Neural Network (CNN) architecture that aggregates local surrounding information for cascaded, fast and accurate detection. The local information serves as context and provides rich cues to distinguish text from background noises. In addition, we designed a novel grouping algorithm on top of detected character graph as well as a text line refinement step. Text line refinement consists of a text line extension module, together with a text line filtering and regression module. Jointly they produce accurate oriented text line bounding box. Experiments show that our method achieved state-of-the-art performance in several benchmark data sets: ICDAR 2003 (IC03), ICDAR 2013 (IC13) and Street View Text (SVT).