Attention Guided Multi-Scale Regression for Scene Text Detection
Ge Huang · 2021 2nd International Conference on Computing and Data Science (CDS) · 2021
Scene text detection plays an important role in terms of its position as the first step of many methods such as text recognition. However, despite of the tremendous improvement made for scene text detection, in the context of this deep learning era, there still remains many challenges such as the variety of text scales and the distinguish difficulty due to the disturbance of background. This paper introduces an attention guided multi-scale regression method for detecting scene text, which achieves promising performance. The pipeline gathers feature maps from two different scales, which enhance its perceptive ability, especially to large and long texts. Moreover, the introduction of attention module further improved the prediction performance in terms of separate the text region from background more accurately. Experiments conducted on ICDAR 2015 and MSRA-TD500 show that comparable results have been achieved by the proposed text detector in terms of accuracy and precision.