WCS-LLENet: A Wavelet-Based Cross-Scale Fusion Network for Low-Light Enhancement of Night Vision Images

Abstract

Night vision images are widely used in night reconnaissance, intelligent surveillance, autonomous driving, and robotic perception, but they often suffer from underexposure, low contrast, texture loss, and non-stationary noise. To address these degradations, this paper proposes a supervised Wavelet Cross-Scale Low-Light Enhancement Network (WCS-LLENet) for night vision image enhancement. The network adopts a dual-branch architecture that combines illumination estimation and detail reconstruction. The illumination branch predicts an illumination gain map for stable linear brightening, while the reconstruction branch learns residual structural compensation to suppress noise and recover textures. A cross-scale feature fusion module is designed to promote bidirectional interaction between high-resolution details and low-resolution semantic features. In addition, discrete wavelet transform and inverse wavelet transform are introduced into skip connections for frequency-aware enhancement. Noise-aware gated fusion is applied at high resolutions, and cross-attention is used at low resolutions to improve edge preservation and semantic alignment. Experiments on LOLv1, LOLv2, SID, and a measured night vision dataset show that WCS-LLENet outperforms representative low-light enhancement methods. The proposed method achieves PSNR/SSIM/LPIPS values of 25.24 dB/0.879/0.091 on LOLv1, 21.58 dB/0.860/0.107 on LOLv2, 24.23 dB/0.775/0.148 on SID, and 19.95 dB/0.805/0.093 on the measured dataset.