WMANet: Wavelet-Based Multi-Scale Attention Network for Low-Light Image Enhancement

Low-light images captured at night often suffer from improper exposure, color distortion, and noise, which degrades the image quality and have a negative influence on subsequent applications. Many existing deep learning-based methods enhance low-light images through spatial domain, which may sacrifice the original image information. In this paper, we put forward a deep learning network for enhancing low-light images based on wavelet transform. We utilize the wavelet transform to divide the image into various frequency scales and then analyze the frequency characteristics of different low-light images in the wavelet domain. The proposed network comprises a low-frequency restoration subnet and high-frequency reconstruction subnet that uses an optimal coefficient of wavelet decomposition to construct a frequency pyramid. Furthermore, we utilized different attention mechanisms to extract frequency information from different images, gradually restoring the brightness information and details of low-light images. Additionally, we utilized a self-constructed multi-scale exposure low-light image dataset for training. Numerous experiments on publicly accessible datasets and our established dataset show that the proposed approach quantitatively and qualitatively surpasses state-of-the-art approaches, particularly for real and complex low-light scenarios. Furthermore, our method produces better visual effects than others and performs well in real-time and real-word downstream vision tasks.