Increasing the Accuracy of Facial Recognition Systems in Poor Lighting Conditions with Adversarial Generative Networks
Increasing the Accuracy of Facial Recognition Systems in Poor Lighting Conditions with Adversarial Generative Networks
Zahra rohani1
1) Instructor, Computer Department, Shirvan Branch, Islamic Azad University, Shirvan, Iran
IOI Code
:
AAHD-ABFHEE
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Face recognition, as one of the most important biometric technologies, has made significant progress in the last decade. However, the performance of these systems in poor lighting conditions such as low light, strong backlight, and non-uniform lighting significantly decreases. This research investigates the role of adversarial generative networks in improving the quality of facial images and increasing the recognition accuracy in challenging lighting conditions. In this paper, first, the basic challenges of facial recognition systems in different lighting conditions are analyzed, and then different GAN architectures including IL-GAN, EnlightenGAN, and advanced attention-based models are investigated. Experimental results show that GAN-based models improve the recognition accuracy from 62% in normal lighting conditions to 89% in challenging lighting conditions by improving the peak signal-to-noise ratio by up to 3.37 dB and the structural similarity index by up to 0.0579. Also, the proposed method based on light-ignoring learning has shown superior performance with an average accuracy of 94.2% in multi-light dataset. This research shows that the use of attention mechanisms and dual generator-cleaner architectures can effectively manage the challenge of poor lighting in face recognition systems.
Keywords : face recognition adversarial generative networks poor lighting conditions image quality enhancement deep learning