FMGDN: Flexible Multi-Grained Dilation Network Empowered Multimedia Image Inpainting for Electronic Consumer

Various consumer electronics have become indispensable for our daily lives, and they rely heavily on visual content, such as image, to deliver immersive experiences to users. However, consumer electronic devices may encounter issues like image corruption, noise interference, or object occlusion, and images in consumer electronics may be subjected to various imperfections, which can be addressed by image inpainting techniques to enhance the visual experience. In this work, we propose a Flexible Multi-Grained Dilation Network (FMGDN) to capture multi-grained information and complete hole regions with semantically and visually plausible contents. Specifically, we design a Multi-Grained Residual (MGR) block with hybrid parallel branches to extract hierarchical features. The branches with different dilation rates use co-prime principle to avoid gridding artifacts. The multi-grained features from different branches are fused with adaptive weights to enforce deeper semantic context. Further, a Channel Adaptive Shuffling (CAS) block is designed to effectively decode high-level features back to image space, shuffling the inner channels with learning-capable filters and adaptive channel weights. In particular, benefitting from its concise and flexible architecture, FMGDN can be easily configured to adapt to various kinds of consumer electronics. Experiments on three real-world datasets demonstrate the effectiveness of FMGDN, outperforming state-of-the-art methods.