sattyss/dlp26t2-nppe3-denoisesr

DLP26T2 NPPE-3 — Low-Light Denoising + 4× Super-Resolution

CNN-based image restoration model for low-light noisy image denoising and 4× super-resolution.

Model

• —Architecture: Residual CNN + PixelShuffle
• —Residual blocks: 16
• —Feature channels: 64
• —Upscaling factor: 4×
• —Input channels: 3
• —Output channels: 3

Training

• —Training pairs: 1105
• —Validation pairs: 267
• —LR patch: 64×64
• —HR patch: 256×256
• —Batch size: 8
• —Epochs: 30
• —Optimizer: AdamW
• —Initial learning rate: 2e-4
• —Weight decay: 1e-4
• —Loss: 0.7 L1 + 0.3 MSE

Validation

Best validation PSNR:

38.731 dB

Input / Output

Input: 256×160 low-light noisy RGB image

Output: 1024×640 restored RGB image

Files

• —model.pth — trained model weights
• —config.json — model and training configuration