最近在研究文字图像的超分辨率,下载了几篇论文,准备略读一下,每篇文章写个笔记做记录。
主要贡献:
This paper studies CNN-based text image SR methods tailored for OCR, and our contributions can be summarized as follows. First, we propose a new loss function when training CNN for text image SR. The new loss function is inspired by the intuition that high-frequency image details play a more important role in OCR, and thus puts more weights on edge regions. Second, we propose a simple yet effective image padding method to refine the image boundaries during CNN- based SR. Third, we conduct model combination to further improve the performance.
Experimental results show that our proposed method achieves an OCR accuracy of as high as 78.10% on the ICDAR 2015 TextSR dataset, which exceeds the state-of-the-art re