Attention-based Feature Weight Fusion Network for Scene Text Detection

Abstract

Due to the diversity of scene texts and background interference, designing an effective and accurate text detector remains challenging. To address this issue, we propose a scene text detection network named as CWDNet, that combines attention mechanisms and weight branches. Firstly, the Coordinate Attention (CA) mechanism is introduced into the residual blocks of the ResNet network to enhance feature extraction. Secondly, a weight branch fusion module is proposed to dynamically adjust the significance of features at different scales. Finally, we conducted experiments on benchmark datasets of ICDAR2015, MSRA-TD500, and ICDAR2017-MLT, the CWDNet achieves F-measure of 85.8%, 84.6%, and 76.3% respectively, demonstrating strong robustness and competitiveness.