Understanding and improving zero-reference deep curve estimation for low-light image enhancement

Jiahao Wu,Dandan Zhan,Zhi Jin

DOI: https://doi.org/10.1007/s10489-024-05534-7

IF: 5.3

2024-05-30

Applied Intelligence

Abstract:Zero-Reference Deep Curve Estimation (Zero-DCE) pioneers a new idea for Low-Light Image Enhancement (LLIE), which is to formulate LLIE as a task of image-specific curve estimation with a deep network. Despite its success, the underlying mechanisms of Zero-DCE still remain under-explored, which prevents the further development of curve estimation methods for LLIE. In this paper, we take a step further in understanding Zero-DCE and provide an in-depth analysis from the perspective of the curve formula design and the loss function balance. Inspired by our analysis, we make effective modifications to Zero-DCE in terms of the curve formula, the curve estimation network and the loss terms, and propose Zero-Reference Exposure Adjusting Curve Estimation (Zero-EACE). A novel curve formula named EAC, a novel curve estimation network named HGNet, and a novel loss function named HE Loss are proposed. Extensive experimental results show that the proposed Zero-EACE achieves comparable performance to state-of-the-art methods both qualitatively and quantitatively. Moreover, experimental results on multiple exposure images demonstrate the capability of our method to simultaneously tackle over- and under-exposure correction, which expands the practical application scenarios of unsupervised curve estimation-based LLIE methods.

computer science, artificial intelligence