Retinex-LTNet: Low-Light Historical Tibetan Document Image Enhancement Based on Improved Retinex-Net

Abstract

Smoothing the background and highlighting the foreground text of document images not only improve the reading experience but also benefits the subsequent document analysis and recognition algorithms. However, most of the past low-light image enhancement methods show less effectiveness while dealing with low-light Tibetan document images. In this paper, we propose an improved Retinex-Net image enhancement network framework for low-light historical Tibetan document images. First of all, to effectively obtain the texture features of the fonts in the document image, we improved the Enhance-Net in Retinex-Net, and introduced the Residual Layer and Dense Connection Block combination block based on the encoder - decoder, so-called Residual Dense Connection Block (RDC Block); Secondly, we constructed a data set containing low/normal-light historical Tibetan document image pairs for network training. The experimental results show that the network not only enhances the image light, but also makes the background and text contrast more vivid. In addition, compared with the original method, the proposed method improves the PSNR and SSIM by 8.1% and 5.8% respectively, reduces the MSE by 21.2%, and achieves better results compared with similar algorithms.