Scene Text Detection with Cascaded Multidimensional Attention

Abstract

Over the past years, scene text detection based on a segmentation network has progressed substantially due to its pixel-level description, which is more suitable for detecting long text and curved text. However, limited by the scale robustness and feature representation ability, most existing segmentation-based scene text detectors may need help to handle more complex forms of text, which is common in the real world. In this paper, to tackle this problem, we propose a cascaded module, termed CMAModule, based on the attention mechanism to improve the feature representation capability of the model, which integrates a series of the basic module to augment the feature map. Our proposed CMANet, obtained higher recall and precision on two benchmarks.