Low-Light Image Enhancement with Attention and Multi-level Feature Fusion

Low-light image enhancement has made impressive progress with convolutional neural networks (CNNs). However, most existing CNNs-based networks ignore the importance of feature channels and multi-level features. To address these issues, we propose a novel low-light image enhancement network. First, we establish Feature Extraction Block (FEB) to extract features and Feature Fusion Block (FFB) to fuse multi-level features. Then, we adopt a compact channel attention module to re-define the channel importance of input features at the beginning of each Feature Extraction Block (FEB) and Feature Fusion Block (FFB). Besides, we adopt two types of low-light image datasets for training, concluding synthetic images and real-world images. Experiments show that our new network makes competitive progress for low-light image enhancement compared with the state-of-the-art methods.