Noise-Resilient Masked Face Detection Using Quantized DnCNN and YOLO

Abstract

This study presents a noise-resilient masked-face detection framework optimized for the NVIDIA Jetson AGX Orin, which improves detection precision by approximately 30% under severe Gaussian noise (variance 0.10) while reducing denoising latency by over 42% and increasing end-to-end throughput by more than 30%. The proposed system integrates a lightweight DnCNN-based denoising stage with the YOLOv11 detector, employing Quantize-Dequantize (QDQ)-based INT8 post-training quantization and a parallel CPU-GPU execution pipeline to maximize edge efficiency. The experimental results demonstrate that denoising preprocessing substantially restores detection accuracy under low signal quality. Furthermore, comparative evaluations confirm that 8-bit quantization achieves a favorable accuracy-efficiency trade-off with only minor precision degradation relative to 16-bit inference, proving the framework's robustness and practicality for real-time, resource-constrained edge AI applications.