Abhirudraa/nppe3-edsr-direct-128

NPPE3 StrongLuminanceSR

A custom 4× image super-resolution model trained for the DLP-26T2-NPPE3 Kaggle competition.

Model

• Architecture: StrongLuminanceSR
• RCAB blocks: 24
• Feature channels: 64
• Input: RGB LR image
• Output: 1-channel luminance (Y)
• Upscaling: 4×
• Upsampling: PixelShuffle
• Residual learning over bicubic RGB → Y baseline

Training

• Patch size: 128
• Loss: 0.8 MSE + 0.2 L1
• Best epoch: 31
• Validation PSNR: 39.387575 dB

Competition Result

Kaggle leaderboard score:

39.17133

Files

• model.pth — trained model weights
• architecture.py — exact PyTorch architecture
• config.json — model configuration

RGB to Y

The model uses:

Y = 0.299R + 0.587G + 0.114B

The network directly predicts the high-resolution luminance channel.

Loading

import torch
from architecture import StrongLuminanceSR

model = StrongLuminanceSR(
    channels=64,
    num_blocks=24,
    scale=4
)

state_dict = torch.load(
    "model.pth",
    map_location="cpu"
)

model.load_state_dict(state_dict)
model.eval()