YOU ONLY NEED 80K PARAMETERS TO ENHANCE IMAGE: LEARNING PERIODIC FEATURES FOR IMAGE ENHANCEMENT

Benefiting from the promising performance of CNNs models for high-level vision tasks, these networks have been extensively adopted to image enhancement tasks. However, recent methods have complex architecture resulting in poor generalization and high computational cost. Their activation functions are originally designed for other vision tasks. In this work, we present a lightweight network to learn periodic features (LPF) using the proposed wave presentation. Specifically, to better capture implicit feature representations, we represent features as signals with three parts: Cosine Wave Map (CWM), Sine Wave Map (SWM) and Direct Current Map (DCM). Thus, we formulate the image enhancement task as a signal modulation problem. Inspired by the Fourier transform, we build the Fourier Enhancement Module (FEM) that allows for efficient and scalable spatial mixing of local and non-local contents and dynamically learns the interaction between waves to enhance the images. LPF with only 80k parameters achieves better quantitative and qualitative results compared with SOTA methods on four image enhancement datasets.