An Auto-Encoding Network for Binary Inscription Character Inpainting based on U-Net and Self-Attention

Many inscriptions have suffered substantial damage as a result of artificial or natural causes, necessitating the use of computers for their repair and preservation. The character reconstruction method, however, is not sufficiently improved by the character inpainting models now in use. The traditional deconvolutional up-sampling module can only extract local spatial points from low-resolution feature maps to produce high-resolution details. To solve this problem, we design Binary Auto-Encoder (BAE) by introducing the self-attention mechanism in the U-Net decoder. The self-attention module calculates the feature weighting at each place to determine responsiveness and thereby computes the attention vector at a very cheap cost. Additionally, we use a spectrum normalization strategy to enhance the decoder’s up-sampling procedure. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of BAE improve by 1.9 and 0.023, respectively, when compared to the baseline model, according to the results of our pairwise comparison of various inpainting models using the binary inscription dataset. The Handwritten Chinese Character Inpainting Generative Adversarial Network (HCCI-GAN) is then compared using the dataset of handwritten Chi-nese characters, and the findings demonstrate that the BAE has a cleverer and simpler architecture.