ScholarWorks@동국대학교

초록

Screenshot demoiréing aims to remove moiré artifacts from images of digital screens. Whereas conventional single-objective learning-based algorithms optimize demoiréing networks by minimizing the distance between the enhanced and ground-truth images, contrastive learning (CL) improves the performance by adding constraints that maximize the difference between enhanced images and negative samples, i.e., degraded input images. However, the complexity of moiré artifacts creates a disparity, leading to under-constrained optimization. Specifically, maximizing the differences with degraded input images is ineffective because they are already substantially distant. In this work, we propose a novel diversity and structure-aware contrastive learning (DiSCo-Learn) strategy that enhances the constraints by generating auxiliary negative samples for screenshot demoiréing. Specifically, we first develop a high-resolution screenshot demoiréing network (HiReSD-Net) that constructs a feature pyramid to handle moiré artifacts across a broad spectrum of frequencies. Then, based on a moiré generation model, we decompose the input moiré images into clean and moiré components, which are subsequently used to generate auxiliary negative samples for DiSCo-Learn. Unlike conventional CL, DiSCo-Learn uses these auxiliary negative samples, enabling HiReSD-Net to exploit the diversity of moiré artifacts and the structure of clean images. This pushes the enhanced images closer to the ground truth, thereby improving the demoiréing performance. Experimental results on real-world, high-resolution datasets demonstrate that the proposed algorithm significantly outperforms state-of-the-art approaches, both quantitatively and qualitatively. © 1991-2012 IEEE.

키워드