Low-Light Image Enhancement Network with Adaptive Feature Fusion and Exposure Dynamic Control

Images captured in military, medical, and other fields often occur at night and are inherently dim, necessitating enhancement. The advent of deep learning has led to the development of several advanced low-light image enhancement techniques. Notably, the self-calibrated illumination (SCI) network stands out due to its distinctive structure and calibration mechanism. This paper introduces a low-light image enhancement network leveraging adaptive feature fusion and exposure control (AFESCI), inspired by the SCI model. We propose an exposure dynamic adjustment network that effectively mitigates uneven illumination using channel-level statistical features and adaptive weighting. Additionally, we present an adaptive local and global feature fusion network that utilizes depthwise separable convolutions to extract regional variance features and establishes cross-region dependencies through multi-head attention. A dynamic gating mechanism is also incorporated to adaptively fuse features from both branches, preventing over-enhancement and distortion. Furthermore, a joint loss function is devised to address brightness fidelity, smoothness, and noise suppression comprehensively. Experimental results highlight the method's qualitative and quantitative advantages.