DarkGAN-Enhanced Low-Light Detection and Localization of Low-Voltage Electricity Meters

This paper proposes LAA-GAN (Luminosity-Aware Attention Generative Adversarial Network for Detection), a lowlight electricity meter detection method designed to improve detection accuracy and real-time performance on edge devices. First, in the low-light enhancement stage, a U-Net-based generator is adopted, incorporating a Luminosity-Aware Attention (LAA) mechanism in the skip connections. This mechanism adaptively assigns higher weights to low-light regions, enhancing detail preservation while integrating multi-scale feature fusion to improve the quality and stability of enhanced images. Next, in the object detection stage, the lightweight YOLOv8-Tiny model is employed for electricity meter localization, ensuring real-time performance. Additionally, GAN training is used to optimize the enhancement process, generating images with a more natural brightness distribution and richer details, ultimately improving the robustness of object detection. Experimental results demonstrate that LAA-GAN effectively enhances electricity meter detection accuracy in various low-light conditions while maintaining low computational complexity, making it suitable for real-time deployment on edge devices.