Wavelet-based enhancement network for low-light image

Abstract

Low-light images are key challenges for high-level vision tasks, often leading to failures in intelligent systems. To achieve more robust low-light enhancement and gain improvement for downstream segmentation task, in this paper we propose a wavelet-based enhancement network (WENet) that combines convolution layer and Transformer block. The wavelet transform separates different frequency components from the multi-scale transformation of the signal. We propose a wavelet calibrate layer (WCL), which converts the feature to the wavelet domain and distributes it to the corresponding area through multiple calibration filters, and restores details of the image. Recognizing that noise amplification occurs concurrently with wavelet learning, we build a contrast adjustment layer (CAL), which refines the contrast primarily through shift operations. WENet has achieved superior performance on the LOL, LOLv2 and MIT-Adobe FiveK datasets for enjoyable visual experience, reaching 22.34 and 0.814 on PSNR and SSIM respectively. We trained WENet and segmentation model by end-to-end in the dark scene of ACDC dataset and achieved advanced effect, which is robust for low-light scenes.