WATNet: A Wavelet-Aware Lightweight Hybrid Model for Fast Low-Light Enhancement

Low-light enhancement task is an essential component of computer low-level visual tasks, which involves processing images captured under dim lighting conditions to make them appear as if they were taken under normal illumination. Currently, deep neural networks have become the mainstream approach for image processing. However, recent works have devoted considerable efforts to designing high-performance models, which often come with high computational complexity and inference time, making real-time processing unfeasible. We observed that some convolutional methods are due to the need for deep layers which results in a large number of parameters. Moreover, enhancing details and removing noise in low-light images remains an open challenge. In order to solve the above problems, we propose a lightweight baseline that combines CNN and sparse grid attention transformer blocks to enable the model to capture a global receptive field at an early stage. Specifically, we propose a High-Frequency Wavelet-aware Block(HFWB) that focuses on processing high-frequency information in the wavelet domain to refine details and suppress noise. With a processing time of only 10.6ms, the performance of our model outperforms that of the current state-of-the-art lightweight models on benchmark low-light datasets. Compared to state-of-the-art models in the LOL dataset, our model achieves a reduction in inference time of over 90% and requires only about 1% of the FLOPS.