A Learnable Frequency Gated DCT Denoising Network

摘要

Most DCT based denoising still follows fixed rules, like hard or soft thresholds, keep only low frequency parts, or simple sub band Wiener filters. On real photos these rules usually trade detail for smoothness, or the other way around. We build a learnable frequency gated denoising network. On top of differentiable orthogonal DCT and IDCT, the model learns masks and gains for high and low frequency channels, and it adds a light low resolution global attention module to catch long range context. With an image level residual skip, the reconstructions look clean and sharp. In our small setup the method lifts PSNR and SSIM a lot compared to a pure DCT baseline, from 11.89 dB and 0.2915 to 25.05 dB and 0.945. In short, we turn a rule driven DCT pipeline into a learnable one, so we keep the frequency prior and still get the robustness of modern learning.