Low Light Image Enhancement using U-Net

This project explores improving image quality in low-light environments using deep learning, with a focus on preserving structural details while enhancing visibility. The objective was to design a model that could jointly perform denoising and exposure correction without introducing artifacts.

I implemented a hybrid U-Net architecture in PyTorch and trained it on the LOLv1 dataset, benchmarking its performance against standard low-light enhancement methods. The model was designed to separate structural and illumination components, enabling more controlled enhancement and reducing common issues such as over-smoothing and color distortion.

The resulting model achieved a PSNR of 19.37 dB and consistently produced visually enhanced images with reduced noise, corrected exposure, and minimal color cast. This project strengthened my understanding of designing task-specific CNN architectures and evaluating image restoration quality both quantitatively and qualitatively.