DFF-Net: Deep Feature Fusion Network for low-light image enhancement

Abstract

Low-light image enhancement methods are designed to improve brightness, recover texture details, restore color fidelity and suppress noise in images captured in low-light environments. Although many low-light image enhancement methods have been proposed, existing methods still face two limitations: (1) the inability to achieve all these objectives at the same time; and (2) heavy reliance on supervised methods that limits practical applicability in real-world scenarios. To overcome these challenges, we propose a Deep Feature Fusion Network (DFF-Net) for low-light image enhancement which builds upon Zero-DCE’s light-enhancement curve. The network is trained without requiring any paired datasets through a set of carefully designed non-reference loss functions. Furthermore, we develop a Fast Deep-level Residual Block (FDRB) to strengthen DFF-Net’s performance, which demonstrates superior performance in both feature extraction and computational efficiency. Comprehensive quantitative and qualitative experiments demonstrate that DFF-Net achieves superior performance in both subjective visual quality and downstream computer vision tasks. In low-light image enhancement experiments, DFF-Net achieves either optimal or sub-optimal metrics across all six public datasets compared to other unsupervised methods. And in low-light object detection experiments, DFF-Net achieves maximum scores in four key metrics on the ExDark dataset: P at 83.3%, F1 at 72.8%, mAP50 at 74.9%, and mAP50-95 at 48.9%. Code is available at https://github.com/WangL0ngTa0/DFF-Net.