LumiNet: A Joint Single-Stage Compact CNN for Low-Light Image Enhancement and Denoising

Abstract

Low-light image restoration remains a challenging problem due to poor illumination, colour distortion, structural detail loss, and amplified sensor noise. Existing restoration methods often rely on multi-stage enhancement and denoising pipelines that increase computational complexity and introduce error propagation during sequential processing. This paper presents LumiNet, a lightweight single-stage convolutional neural network for joint low-light image enhancement and denoising. The proposed framework introduces a Tri-Map Joint Restoration Block (TM-JRB) that estimates illumination, noise, and structure guidance maps to enable adaptive feature reconstruction across degraded image regions. Within the TM-JRB, a Wavelet-Inspired Brightness Refinement Block improves low-frequency illumination estimation, while an FFT-based frequency-aware refinement module refines amplitude while preserving phase information. Furthermore, Signal-to-Noise Ratio Guided Skip Fusion (SNR-GSF) selectively transfers reliable encoder features to improve detail preservation while suppressing noise. Experimental evaluations on multiple enhancement, denoising, and joint restoration benchmarks demonstrate that LumiNet achieves competitive restoration performance while maintaining computational efficiency with only 1.49 million trainable parameters. The proposed framework achieves 21.13 dB PSNR and 0.84 SSIM on the LOLv2-Real dataset and 35.23 dB PSNR on the SIDD benchmark. Overall, LumiNet provides an effective balance between enhancement quality, denoising performance, structural preservation, and computational efficiency for real-world lowlight image restoration.