This project implements an Invertible Rescaling Network (IRN) for single-image denoising. The IRN framework learns a bijective mapping between noisy and clean image domains, preserving information through invertible transformations. This approach enables lossless feature reconstruction and efficient memory utilization. 

Highlights 

• Employs invertible blocks to maintain exact information flow. 

• Supports bidirectional learning between noisy and clean domains. 

• Achieves strong denoising performance with reduced computational overhead.

This work applies the U2-Net architecture, originally designed for salient object detection, to the image denoising task. By leveraging nested U-structures, the model effectively captures both global context and fine details for accurate noise removal.

Highlights 

• Uses nested U-blocks to enhance feature representation. 

• Retains image edges and fine textures during reconstruction. 

• Demonstrates robustness across multiple noise levels and datasets.