Reproduce “FreqNet: A Frequency-domain Image Super-Resolution Network with Dicrete Cosine Transform” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Runyuan Cai, Yue Ding, Hongtao Lu
FreqNet, developed by researchers at Shanghai Jiao Tong University, approaches Single Image Super-Resolution by explicitly operating in the frequency domain using Discrete Cosine Transform. This method enables direct manipulation of frequency components, yielding reconstructions with superior high-frequency fidelity and competitive PSNR values while mitigating hallucinated artifacts.
Key insights
- Explicitly operating in the frequency domain via Discrete Cosine Transform (DCT) allows direct manipulation and reconstruction of high-frequency components, enabling better control over detail generation and fidelity.
- A dual-path network architecture, combining a Spatial Extraction Network (SEN) with a Frequency Reconstruction Network (FRN), effectively processes both spatial and frequency domain information.
- Specialized residual groups (Deformable Residual Groups for spatial and Depth-wise Residual Groups for frequency) and a perceptually weighted frequency-domain loss function enhance feature learning and reconstruction accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.