A robust GAN-based model for low-light image enhancement and human detection on mobile devices

Detecting people in low-light environments remains difficult for most artificial intelligence systems. Changes in illumination-caused by time of day, weather, or artificial lighting-often reduce the reliability of visual recognition algorithms. When light levels drop, images lose detail and gain noise, making it harder to identify people accurately in situations such as nighttime surveillance or security monitoring, where clear visibility is crucial. To address this issue, we propose an illumination enhancement model based on Generative Adversarial Networks (GANs) that reconstructs well-illuminated images from low-light inputs using the ratio-log concept. The enhanced outputs reached 25.15 PSNR and 0.909 SSIM on the LOLv1 dataset, and 28.04 PSNR and 0.905 SSIM on a surveillance-oriented dataset, outperforming well-known enhancement methods. Moreover, a YOLO-based human detection test confirmed that brighter images increased detection confidence from 0.44 to 0.78 on average and revealed people invisible in the original frames. Additionally, the GAN model was integrated into an Android application built under clean architecture principles, enabling real-time enhancement directly on the device. Finally, the Usability evaluations with IT professionals yielded positive feedback, confirming the model’s effectiveness and accessibility for real-world use.