核心概念
AADNet proposes a novel lightweight architecture for high-resolution image demoiréing that effectively works across different frequency bands and generalizes well to unseen datasets.
摘要
Abstract:
Moiré patterns in photographs degrade image quality.
Existing methods struggle with dynamic textures and variations in moiré patterns.
AADNet is proposed for effective demoiréing.
Introduction:
Digital sensors simplify image acquisition but introduce moiré patterns.
Benchmark study reveals limitations of existing methods.
AADNet aims to address these challenges.
Related Work:
Researchers tackle Moiré patterns from synthetic to real-world challenges.
Existing methods fall short in handling wider moiré patterns in 4K images.
Methodology:
AADNet utilizes an attention-aware approach for finer details in document demoiréing.
SAM Blocks and Global Attention blocks are strategically embedded within the decoder levels.
Experiments:
Conducted on the UHDM dataset to improve text details in demoiréd document images.
Evaluation metrics include PSNR, SSIM, and LPIPS.
Results:
Quantitative evaluation shows state-of-the-art performance on the UHDM dataset.
Qualitative comparison demonstrates better perceptual results compared to other methods.
Conclusion:
AADNet achieves accurate and robust moiré removal on ultra-high-definition images.
Proposed architecture incorporates Focal Frequency Loss and Attention-aware blocks for improved performance.
統計資料
AADNet exceeds multi-stage high-resolution method FHDe2Net, 3 dB in terms of PSNR while being 300× faster (5.620s vs 0.017s) in the UHDM dataset.