LAE-Net: A locally-adaptive embedding network for low-light image enhancement

Abstract

In the low-light enhancement task, one of the major challenges lies in how to balance the image enhancement properties of light intensity, detail presentation and color fidelity. In natural scenes, the multi-distribution of frequency and illumination characteristics in the spatial domain makes the balance more difficult. To solve this problem, we propose a Locally-Adaptive Embedding Network, namely LAE-Net, to realize high-quality low-light image enhancement with locally-adaptive kernel selection and feature adaptation for multi-distribution issues. Specifically, for the frequency multi-distribution, we rethink the spatial-frequency characteristic of human eyes, experimentally explore the relationship among the receptive field size, the image spatial frequency and the light enhancement properties, and propose an Entropy-Inspired Kernel-Selection Convolution, where each neuron can adaptively adjust the receptive field size according to its spatial frequency characterized by information entropy. For the illumination multi-distribution, we propose an Illumination Attentive Transfer subnet, where the neurons can simultaneously sense global consistency and local details, and accordingly hint where to focus the efforts on, thereby adjusting the refined features. Extensive experiments with ablation analysis show the effectiveness of our method and the proposed method outperforms many related state-of-the-art techniques on four benchmark datasets: MEF, LIME, NPE and DICM.