Bidirectional Regression for Arbitrary-Shaped Text Detection

Arbitrary-shaped text detection has recently attracted increasing interests and witnessed rapid development with the popularity of deep learning algorithms. Nevertheless, existing approaches often obtain inaccurate results, mainly due to relatively weak ability utilize context information inappropriate choice offset references. This paper presents a novel instance expression which integrates both foreground background into pipeline, naturally uses pixels near boundaries as starts. Besides, corresponding post-processing algorithm is also designed sequentially combine four prediction results reconstruct accurately. We evaluate our method on several challenging scene benchmarks, including curved multi-oriented datasets. Experimental demonstrate that proposed approach obtains superior or competitive performance compared other state-of-the-art methods, e.g., 83.4\(\%\) F-score for Total-Text, 82.4\(\%\) MSRA-TD500, etc.