Lightweight Deep Learning for Real-Time Edge Object Detection in Embedded Electronic Systems

Abstract

Edge computing in electronic and information systems requires object detection algorithms that balance high accuracy with low computational complexity ${ }^{[1]}$. Traditional deep convolutional neural networks (CNNs) often fail to meet real-time processing constraints on resource-constrained embedded platforms due to excessive parameter size and memory footprint ${ }^{[2,3]}$. This paper proposes a lightweight neural network architecture, MobileDetect-Net, designed for real-time edge processing. Specifically, the proposed Lite-FPN module introduces a lightweight multi-scale fusion mechanism that minimizes parameter redundancy, while the INT8 quantization deployment maximizes hardware arithmetic throughput without significant precision degradation. Experimental evaluations conducted on an ARM Cortex-A72 embedded hardware platform demonstrate that MobileDetect-Net reduces total parameters by 68.1% compared to standard YOLOv5s ${ }^{[1,3]}$, while achieving a mean average precision (mAP@0.5) of 84.2% at 38.5 frames per second (FPS) ${ }^{[6]}$. Ablation results further show that the complete configuration attains 78.6% mAP under low light and 72.1% mAP under motion blur. These results support the effectiveness of the proposed model for real-time electronic information processing on edge devices.