UmmiShifa/lowlight-denoise-sr

Low-Light Denoising + 4x Super-Resolution

Fine-tuned EDSR for joint low-light image denoising and 4x super-resolution.

Task

Denoise low-light, noisy images affected by poor visibility and upscale them 4x, producing clean, high-resolution output.

Approach

• Base model: EDSR (eugenesiow/edsr)
• Fine-tuning: trained end-to-end directly on noisy low-resolution images paired with clean high-resolution ground truth, at a low learning rate (5e-5) to adapt the pretrained weights without destroying their existing image priors.
• Data augmentation: random horizontal/vertical flips and 90° rotations.
• Inference: 8x self-ensemble test-time augmentation (predictions averaged over flipped/rotated versions of each input, then aligned back) for a reliable PSNR boost with no extra training.