IoU-Related Arbitrary Shape Text Scoring Detector

Abstract

As a medium of information transmission, text is widely existed in natural scenes, showing diversity in orientation, scale, font, color and shape. Accurate detection of scene text is a prerequisite for subsequent recognition. Though many previous methods have worked well on horizontal and multi-oriented text detection datasets, detecting arbitrary shape scene text still remains as a challenging problem. To solve the problem, this paper proposes an arbitrary shape scene text detection method. Based on Mask R-CNN, our method replaces the original ℓ1-smooth loss with the proposed IoU-related loss and adds a text scoring branch to align the confidence score with the text mask IoU to make the model highly relevant to IoU, achieving the goal of improving detection performance by improving IoU directly in a simple but effective way. The proposed method is evaluated on four public datasets: CTW-1500, Total-Text, ICDAR2015 and ICDAR2017-RCTW. For curved text detection datasets CTW-1500 and Total-Text, we have reached 79.2% and 81.1% H-mean respectively, showing that the proposed method has achieved competitive performance in arbitrary scene text detection.

Keywords

• Scene text detection
• arbitrary shape
• computer vision
• intersection-over-union
• semantic segmentation