基于空间注意力的文本超分辨率重建算法

Text super resolution network based on spatial attention

自然场景的文本识别在深度学习的推动下发展迅速,但是大多数文本识别算法都无法在低分辨率条件下使用,为了解决这个困境,本研究将超分辨率重建技术作为预处理手段加入到文本识别中,并提出一种针对文本识别领域的超分辨率重建网络,利用像素级注意力机制使超分辨率重建网络专注文本相关信息,同时利用文本识别网络的先验信息构建感知损失,辅助超分辨率识别网络的训练,将低分辨率图片的识别率从66.38%提升到78.99%.

The development of sence text recognition is rapid,but most text recognition algorithm has poor performance under condition of low resolution.To solve this problem,super-resolution algorithm is introduced as a preprocess method.In this paper,we proposed a text super resolution network based on spatial attention,meanwhile,training strategy like text prior perceptual loss and edge loss was implemented to improve the preformace on text image.The accuracy of text recognition increased from 66.38% to 77.89% with the help of super resolution network.