LCA-Net: A Context-Aware Lightweight Network for Low-Illumination Image Enhancement
Zhenghao Shi, Manyu Wang, Wenqi Ren
Abstract
Low light image enhancement has made great progress owing to powerful deep representation learning. However, due to the introduction of a large number of parameters, the computation cost of a deep learning based method increases dramatically, which poses great challenges for computational resources, and results in difficult to deploy the method on conventional imaging instruments. To address this problems, this paper proposes a context-aware lightweight model for low illumination image enhancement (Simply termed as LCA-Net). In which, to achieve a brightness balance between the overall image and local details, a context based modeling technique for modeling the non-local dependency of image pixels is developed, and a contrast enhancement model is constructed. To avoid overexposure, Gaussian function is employed to re-establish the light distribution mapping between different image channels. Finally, to avoid the loss of high-frequency details caused by smoothing operations, the fast Fourier transform (FFT) loss is introduced to constrain the network training. Considerable evaluations show that the proposed method performs favorably against the compared counterparts. Moreover, extensive experiments on public datasets have shown that the evaluation metrics and performance of LCA-Net are superior to current state-of-the-art methods.