Abstract

Low-light video capture often suffers from poor visibility, low contrast, and excessive noise, which significantly degrades visual quality and hinders subsequent video analysis tasks. Traditional enhancement techniques struggled to balance noise suppression and detail preservation. Existing methods often either over-smooth the video frames, leading to loss of important textures, or fail to remove noise effectively, resulting in distorted low-light outputs. This challenge is compounded in dynamic video sequences where temporal consistency is critical. This study proposed a deep hybrid framework that combined Convolutional Neural Networks (CNNs) for feature extraction and Generative Adversarial Networks (GANs) for realistic video reconstruction. Initially, input low-light frames were preprocessed using histogram equalization and denoising filters to normalize brightness and reduce high-frequency noise. The CNN component extracted spatial features and enhanced structural details, while the GAN module employed an adversarial loss to generate visually plausible frames. Temporal consistency was enforced through a frame-recurrent approach that aligned consecutive frames, mitigating flickering and maintaining smooth transitions. The hybrid training leveraged both pixel-level loss for detail retention and perceptual loss to improve visual realism. Experimental evaluation demonstrated that the proposed framework achieved substantial improvements over conventional methods. The Peak Signal-to-Noise Ratio (PSNR) reached 28.5 dB, Structural Similarity Index Measure (SSIM) 0.91, Flicker Index 0.10, and Visual Quality Score 4.7/5, indicating enhanced noise suppression, structural fidelity, and temporal coherence across low-light video sequences.

Authors

S. Selvi¹, Pitty Nagarjuna²
Sethu Institute of Technology, India¹, Indian Institute of Science, Bengaluru, India²

Keywords

Low-Light Video Enhancement, Noise Suppression, Hybrid CNN-GAN, Video Denoising, Temporal Consistency