Arbitrary-Shaped Text Detection With Watershed Segmentation Network

Arbitrary-shaped text detection is an important task in computer vision which has achieved great success in text reading systems. However, there still exist several challenges to be solved. For example, the regression-based text detection methods always fail to fit highly-curved text contours accurately. While the semantic segmentation-based methods could fit any geometries accurately, most of them employ single-size convolution kernels which ignore the diversity of receptive fields. To tackle the above issues, we propose a Watershed Segmentation Network (WSS-Net) which consists of two stages. Firstly, the pixel-level predictions of masks including text regions mask and text kernel regions mask are generated by the semantic segmentation network, in which two strategies i.e., deepening feature fusion and expanding receptive field are designed. Thus, the generated region masks can fit the contours of various-scale text instances better. Secondly, a rebuilding module based on the watershed algorithm is proposed to rebuild text instances according to the original image and two region masks, which is a widely used post-processing technique. Extensive experiments are carried out on Total-Text, CTW1500, and MSRA-TD500 to evaluate the effectiveness of the proposed WSS-Net. The results show that our WSS-Net achieves the highest F-measure score against several state-of-the-arts on three datasets.