Degraded Structure and Hue Guided Auxiliary Learning for low-light image enhancement

Most of existing low-light image enhancement (LLIE) approaches typically require expensive pairs of training data, which presents significant challenges in practice. On the other hand, unsupervised methods, that depend on unpaired data and handcraft prior knowledge, often suffer from problems such as structural blurring, color distortion, and even unpredictable and ineffective enhancement in complex scenarios. To address this issue, we propose a novel unsupervised method, i.e., Degraded Structure and Hue Guided Auxiliary Learning (SHAL-Net). It is designed to be efficient, light-weight and strike a balance between inference speed, low data dependency, and visually pleasing result. SHAL-Net extensively explores degraded features of low-level input from multiple dimensions, which enables self-guided auxiliary learning. Specifically, we present a learnable Structure-contrast Recovery (SCR) module that integrates tightly with our designed Decomposition Net (Decom-Net). Through such cascaded auxiliary learning, Decom-Net and SCR mutually constrain each other. This regularization technique restricts the parameter space of Decom-Net, helping it avoid local optima. Moreover, we also introduce an efficient Enhancement stage in inference, further improving the performance. Extensive experiments and ablation studies are conducted on several public benchmarks, demonstrating the superiority of SHAL-Net over the state-of-the-art unsupervised methods and even some representative supervised approaches. Our code and models are available at: https://github.com/hmx-harry/SHAL-Net.