Abstract
Image restoration, which aims to recover high-quality images from theircorrupted counterparts, often faces the challenge of being an ill-posed problemthat allows multiple solutions for a single input. However, most deep learningbased works simply employ l1 loss to train their network in a deterministicway, resulting in over-smoothed predictions with inferior perceptual quality.In this work, we propose a novel method that shifts the focus from adeterministic pixel-by-pixel comparison to a statistical perspective,emphasizing the learning of distributions rather than individual pixel values.The core idea is to introduce spatial entropy into the loss function to measurethe distribution difference between predictions and targets. To make thisspatial entropy differentiable, we employ kernel density estimation (KDE) toapproximate the probabilities for specific intensity values of each pixel withtheir neighbor areas. Specifically, we equip the entropy with diffusion modelsand aim for superior accuracy and enhanced perceptual quality over l1 basednoise matching loss. In the experiments, we evaluate the proposed method forlow light enhancement on two datasets and the NTIRE challenge 2024. All theseresults illustrate the effectiveness of our statistic-based entropy loss. Codeis available at https://github.com/shermanlian/spatial-entropy-loss.