Enhancing Face Image Inpainting via Low-Parameter Multi-Order Feature Interaction

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

In computer vision, face image inpainting aims to reconstruct the lost or damaged regions, maintain visual realism. While traditional methods struggle with large missing areas or complex textures, deep learning based approaches, despite promising achievements, often come at the cost of substantial computational resources To mitigate these challenges, this paper proposes a low-parameter multi-order feature interaction method. It introduces a shadow module that utilizes low-cost linear transformations to enhance feature extraction accuracy. Additionally, a multi-order aggregation module captures and encodes middle-order features, overlooked by traditional methods, enhancing model robustness and generalization. To further reduce parameters, a fusion moment channel attention module is proposed, which optimizes feature map weighting through cross-channel fusion of multi-order statistical information. Extensive experiments on the CelebA-HQ dataset demonstrate that our method surpasses existing approaches while significantly reducing the number of parameters. our approach achieves a PSNR of 29.58 dB and an SSIM of 0.8956 for a mask ratio of (0.1, 0.6], and the parameter is 9.66 M, highlighting its effectiveness in face image inpainting tasks.

Article activity feed