nppe-3

dlp26t2-nppe3-denoise-sr

Fine-tuned Real-ESRGAN (RRDBNet, x4plus) for joint low-light image denoising and 4x super-resolution, trained for the DLP26T2 NPPE-3 Kaggle competition.

Task

Given a noisy, low-light, low-resolution input image, produce a denoised output at 4x the input resolution.

Base model

• Architecture: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
• Pretrained init: official RealESRGAN_x4plus.pth (https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth)

Fine-tuning

• Data: competition train pairs (gt/ = clean high-res target, train/ = noisy low-res input)
• Loss: L1 (pixel) loss
• Optimizer: Adam (lr=1e-4, betas=(0.9, 0.99)), cosine LR schedule
• Epochs: <filled in automatically by the notebook — see EPOCHS>
• Batch size: 2 x 4 gradient-accumulation steps (effective batch 8)
• Patch size: 64 (LR) / 256 (HR)
• Mixed precision (torch.amp)

Evaluation

• Best validation PSNR: <filled in automatically by the notebook — see best_psnr> dB

Usage

import torch
from basicsr.archs.rrdbnet_arch import RRDBNet

model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
state_dict = torch.load('best_model.pth', map_location='cpu')['params_ema']
model.load_state_dict(state_dict)
model.eval()

# input: float tensor in [0,1], shape (1, 3, H, W)
# output: float tensor in [0,1], shape (1, 3, H*4, W*4)
with torch.no_grad():
    output = model(input_tensor).clamp(0, 1)

Limitations

• Fine-tuned on a small competition-specific dataset for a short training budget (a couple of days); not intended as a general-purpose super-resolution model.
• Validated with a single-scale L1 loss + PSNR — no perceptual/adversarial loss was used in this fine-tune, so outputs may look slightly less "sharp" than a full Real-ESRGAN/GAN-trained checkpoint at the same PSNR.

License

Base model released under BSD-3-Clause by the original Real-ESRGAN authors (xinntao et al.). This fine-tuned checkpoint inherits that license.