On the Use of Attention in Deep Learning Based Denoising Method for Ancient Cham Inscription Images.

Image denoising is one of the most important steps in the document image analysis pipeline thanks to its good effect into the rest of the workflow. However, the noise in historical documents is totally different from the common noise present in other classical problems of image processing. It is particularly the case of the image of Cham inscriptions obtained by the stamping of ancient stele. In this paper, we leverage the advantage of deep learning to adapt with these noisy conditions. The proposed network follows an encoder-decoder structure by combining convolution/deconvolution operators with symmetrical skip connections and residual blocks for improving reconstructed image. Furthermore, global attention fusion is proposed to learn the relevant regions in the image. Our experiments demonstrate the proposed method can’t only remove unwanted parts in the image, but also enhance the visual quality for the Cham inscriptions.