MDFF-Net: Multi-level Dynamic Feature Fusion Network Combined with Deep Supervision Mechanism for Low-Light Image Enhancement

Images captured in low-light environments often suffer from various degradations, and numerous effective solutions have been proposed to address these issues. However, current approaches still have problems such as fuzzy details, distorted colors, and insufficient contrast. To enhance the quality of low-light images, we develop a multi-level dynamic feature fusion network combined with a deep supervision mechanism. More specifically, we propose a multi-scale input encoder (MIE) in the feature extraction part, which aims to help the network perceive the scale differences of objects in low-light images, thereby enabling the network to efficiently restore image details. Furthermore, we develop a Multi-level Dynamic Feature Fusion Module (MDFFM) to dynamically fuse features at different levels to complement each other, which helps the network suppress low-light image noise, capture richer color information, and minimize redundant features during the fusion process. When reconstructing images, we utilize multi-level deep supervision in the network, which can solve the vanishing gradients by adding supervision signals in the side outputs, while guiding the network to achieve image enhancement through a stepwise reconstruction manner. Additionally, the output image's contrast is improved by the Contrast Enhancement Module (CEM). Extensive test results show that our method can improve the visual experience of the results and achieve optimal performance across multiple evaluation metrics.