Low-Light Image Enhancement for Edge-Based Security Surveillance in 6G-IoT Visual Systems

Application areas such as real-time visual analytics over high-bandwidth 6G networks, low-power camera networks in remote or low-light environments and surveillance drones, usually operate under insufficient lighting conditions. The captured images are often of low quality, poor resolution and poor visual clarity, leading to reduced visibility, color distortion, and amplified noise. Existing methods of Low-light image enhancement suffer from low accuracy with compromised reliability, trust and fairness. Inspired by the zero-reference learning paradigm of Zero-DCE++, this work aims to investigate the impact of data pre-processing and augmentation strategies for improving the performance of real-time, mission-critical security systems where low-light surveillance images are used for critical decision making. The proposed method uses FFDNet for denoising, exposure fusion for illumination improvement and data augmentation for bias mitigation and performance optimization through diverse training samples. The method is curated for edge deployment on constrained IoT hardware, with low latency and energy efficient usage in 6G-IoT visual systems. The proposed model is aimed at performance improvement on trust driven visual improvements, reduced distributional bias, and deployment fairness across diverse lighting conditions and scenarios. Comparative analysis demonstrates that with the help of zero-reference deep curve estimation, the proposed, DA-Zero-DCE++, pipeline achieves improved performance as compared to state-of-the-art low-light image enhancement methods. Our best configuration, which combines exposure fusion-based augmentation and mild denoising using FFDNet, achieves an average PSNR of 15.34 dB, SSIM of 0.4869, and MAE of 40.87 on the SICE dataset at 1200 × 900 resolution. For high-level vision applications such as real-time visual analytics over highbandwidth 6G networks, low-power camera networks in remote or low-light environments, the performance is further validated on DarkFace dataset where high average precision at intersection over union of 0.5 is achieved.