Detail-enhanced image inpainting based on discrete wavelet transforms

• The contents and textures of an image to be inpainted are separately generated by a two-parallel-branch network. • A multi-level fusion module is proposed to improve the network capability in semantic understanding. • A spatially discounted mask is designed to evaluate the roles of missing pixels with different importance. Deep-learning-based method has made great breakthroughs in image inpainting by generating visually plausible contents with reasonable semantic meaning. However, existing deep learning methods still suffer from distorted structures or blurry textures. To mitigate this problem, completing semantic structure and enhancing textural details should be considered simultaneously. To this end, we propose a two-parallel-branch completion network, where the first branch fills semantic content in spatial domain, and the second branch helps to generate high-frequency details in wavelet domain. To reconstruct an inpainted image, the output of the first branch is also decomposed by discrete wavelet transform, and the resulting low-frequency wavelet subband is used jointly with the output of the second branch. In addition, for improving the network capability in semantic understanding, a multi-level fusion module (MLFM) is designed in the first branch to enlarge the receptive field. Furthermore, drawing lessons from some traditional exemplar-based inpainting methods, we develop a free-form spatially discounted mask (SD-mask) to assign different importance priorities for the missing pixels based on their positions, enabling our method to handle missing regions with arbitrary shapes. Extensive experiments on several public datasets demonstrate that the proposed approach outperforms current state-of-the-art ones. The codes are public available at https://github.com/media-sec-lab/DWT_Inpainting .