Zero-shot contrast enhancement and denoising network for low-light images

Low-light image enhancement has wide applications. However, contrast improvement and denoising are easily overlooked in existing low-light image enhancement algorithms. Inspired by the technique of zero-shot learning, a zero-shot contrast enhancement and denoising network is proposed to remedy the above disadvantages. First, different hierarchical features are extracted by a multi-scale dense network, where the features in the previous layers can be fully used. This step can obtain richer features from the observed low-light image. Second, a hierarchical feature distillation block, including channel shuffle, contrast attention mechanism and noise attention mechanism, is designed to refine the extracted features. This step contributes to contrast enhancement and denoising. Finally, a mapping network is employed to adjust the brightness, which can map the refined features to the enhanced image in pixel-wise way. The proposed network does not require any reference samples during the training phase, and non-reference loss functions are designed to improve the performance. Subjective and objective experiments demonstrate the superiority of the proposed method in contrast improvement, denoising, brightness enhancement and naturalness preservation.