21f2000709_DLP_26T2_NPPE3

Low-Light Denoising & 4x Super-Resolution

This repository contains fine-tuned weights for joint low-light image denoising and 4x super-resolution.

Architecture

• Model: RRDBNet (imported from basicsr.archs.rrdbnet_arch)
• Parameters: ~16.7 Million

Architecture Configuration

The architecture is configured in config.json with the following parameters:

• scale: 4 (4x upscaling)
• num_in_ch: 3 (RGB input channels)
• num_out_ch: 3 (RGB output channels)
• num_feat: 64 (intermediate feature channels)
• num_block: 23 (Residual-in-Residual Dense Blocks)
• num_grow_ch: 32 (dense growth channels)

Performance

• Baseline (Bicubic): 33.02 dB PSNR
• Fine-Tuned Model: 39+ dB PSNR (>6 dB improvement)

Usage

pip install -r requirements.txt
python inference.py input_lr.png output_hr.png best_model.pth

Or in Python:

from inference import enhance_image

enhance_image("test_lr.png", "output_hr.png", "best_model.pth")

Files in this Repository

• best_model.pth: Fine-tuned model weights checkpoint
• config.json: Architecture configuration and training parameters
• inference.py: Minimal inference script
• requirements.txt: Python package dependencies