AT-Text: Assembling Text Components for Efficient Dense Scene Text Detection

Abstract

Text detection is a prerequisite for text recognition in scene images. Previous segmentation-based methods for detecting scene text have already achieved a promising performance. However, these kinds of approaches may produce spurious text instances, as they usually confuse the boundary of dense text instances, and then infer word/text line instances relying heavily on meticulous heuristic rules. We propose a novel Assembling Text Components (AT-text) that accurately detects dense text in scene images. The AT-text localizes word/text line instances in a bottom-up mechanism by assembling a parsimonious component set. We employ a segmentation model that encodes multi-scale text features, considerably improving the classification accuracy of text/non-text pixels. The text candidate components are finely classified and selected via discriminate segmentation results. This allows the AT-text to efficiently filter out false-positive candidate components, and then to assemble the remaining text components into different text instances. The AT-text works well on multi-oriented and multi-language text without complex post-processing and character-level annotation. Compared with the existing works, it achieves satisfactory results and a considerable balance between precision and recall without a large margin in ICDAR2013 and MSRA-TD 500 public benchmark datasets.