Text Detection Based on Bidirectional Feature Fusion and SA Attention Mechanism

Most of the scene text detection methods can not balance accuracy and speed, so these text detection methods are difficult to apply to the scenarios with real-time requirements. We propose a text detection network that attempts to improve text detection’s accuracy while guaranteeing the speed of text detection. Firstly, we use ResNet with SA attention mechanism as a backbone network for extracting image features. Then, different feature layers are fused by a bidirectional feature fusion mechanism, and a bilinear interpolation method is adopted for upsampling to reduce the feature loss in the feature fusion process. Finally, the post-processing part uses a differentiable binarization module to ensure the generalization and real-time performance of the model. The model has good performance on public dataset, achieving 41 of FPS and 83.2% of F-measure on the ICDAR2015, which verify the model has both good real-time performance and high accuracy.