Low-Light Image Enhancement Using Image-to-Frequency Filter Learning

Abstract

Low-light image (LLI) enhancement techniques have recently demonstrated remarkable progress especially with the use of deep learning (DL) approaches. Yet most existing techniques adopt an image-to-image learning paradigm where DL model architectures are constrained due to latent image feature reconstruction. In this paper, we propose a new LLI enhancement solution titled LLHFNet (Low-light Homomorphic Filtering Network) which performs image-to-frequency filter learning. It is designed independently from custom DL architectures and can be seamlessly coupled with existing feature extractors like ResNet50 and VGG16. We have conducted a large battery of experiments using SICE and Pascal VOC datasets to evaluate LLHFNet’s enhancement quality. Our solution consistently ranks among the best existing image enhancement techniques and is able to robustly handle LLIs and normal-light images (NLIs).