Large Kernel Convolutional Attention Based U-Net Network for Inpainting Oracle Bone Inscription

Abstract

Oracle Bone Inscription (OBI) image inpainting is important for inheriting the long history of Chinese culture. The complexity of OBI characters leads to many unique writing characteristics, which increase the difficulty of restoring irregular structures and multiple types of strokes. In this work, we formulate a large kernel convolutional attention based U-Net framework to restore OBI images. The framework consists of two modified U-Nets arranged in a series, which perform an edge inpainting function and an overall image inpainting function, respectively. The implementations of the two U-Nets are identical and each U-Net is composed of an encoder that downsamples input images, followed by eight large kernel convolutional attention blocks, and a decoder that upsamples the image back to its original size. In addition, an adversarial learning algorithm with local and global discriminative networks is used to train the proposed framework to obtain OBI inpainting results. Compared with state-of-the-art image inpainting methods, our experimental results show that the proposed method can achieve the best inpainting results in OBI images; in particular, the method is also suitable to handle the large-area mask tasks.