Arbitrary-Shaped Text Detection with Gaussian Probability Distance Distribution

Abstract

With the development of semantic segmentation, segmentation-based methods have yielded great success in detecting arbitrary-shaped texts. However, many existing text detection methods use binary discrete distributions to predict shrunk text instances, which cannot generate complete and accurate text bounding boxes. In this paper, we propose an arbitrary-shaped scene text detection method based on predicting Gaussian probability distance map of the complete text region, and this map can retain more text boundary information. Then, the boundary pixels are clustered into high-confidence text centers by a learnable post-processing and false positives are filtered out by pixel-level score maps. We also propose an adaptive channel enhancement module to improve the pixel-level segmentation accuracy. Experiments on three standard datasets, including CTW1500, Total-Text, and MSRA-TD500, demonstrate that the proposed method achieves great robustness and performance. The method obtains an F-measure of S2.S% on CTW1500 and S3.0% on MSRA-TD500.