Arbitrary Scene Text Detection with Bezier Proposal

Abstract

Scene text detection is widely studied in natural language processing since 2016, in which arbitrary scene text detection is always the difficulty. At present, to deal with the problem of how to detect arbitrary shape text, the semantic segmentation-based methods are widely used, but the post-processing and label generation operations are complex. Sparse R-CNN is a novel object detection framework with simple process and high accuracy, which can simplify post-processing by bipartite graph matching loss. Therefore, an arbitrary shape text detect method without any post-process based on Bezier proposal with Sparse R-CNN is proposed. Firstly, the feature pyramid network with attention mechanism is used to extract features, and then the processed features go into the Sparse R-CNN detection head to get the score and coordinates, and finally the detection results are visualized according to the score. The results on ICDAR2015 and CTW1500 datasets show that our method can detect arbitrary text effectively, and our method have higher accuracy and higher speed than other methods.