Scholarlink

<p>Deploying deep learning-based facial recognition on resource-constrained IoT devices presents significant challenges due to high computational demands, memory requirements, and power consumption. This paper presents a novel hybrid lightweight architecture for contactless smart attendance systems, specifically optimized for edge computing environments such as IoT-enabled classrooms. Unlike conventional end-to-end deep learning approaches, our system strategically combines deep neural networks with classical machine learning techniques to achieve real-time performance on low-power hardware while maintaining high accuracy. The proposed system employs a five-stage modular pipeline: (1) Motion detection using a low-cost infrared sensor to activate the camera only when movement is detected, reducing power consumption and unnecessary processing. (2) Face detection through an optimized Viola–Jones algorithm employing Haar features, integral images, and a cascade classifier for efficient identification of face regions under constrained resources. (3) Feature extraction using a lightweight FaceNet model trained with triplet loss to generate compact and discriminative 128-dimensional embedding. (4) Face classification using a Support Vector Machine (SVM) that accurately distinguishes identities based on the extracted embedding while minimizing computational load. (5) Automated attendance management with real-time database synchronization and structured logging for administrative verification. Comprehensive experimental evaluation demonstrates superior performance across multiple benchmark datasets: the proposed Viola-Jones + FaceNet + SVM model achieved 98% accuracy on the LFW dataset and 99.8% on the NUAA dataset, outperforming all compared approaches. Most significantly, the system achieved 100% accuracy on our real-world classroom database, demonstrating exceptional practical reliability in authentic deployment scenarios. Implementation on Raspberry Pi 4B achieves an average inference time of 0.68&#xa0;s per face with only 512&#xa0;MB memory footprint, representing substantial improvements in computational efficiency compared to end-to-end deep learning alternatives. The system successfully handles real classroom conditions including varying illumination, multiple face angles, and practical deployment constraints. This work demonstrates that carefully designed hybrid architectures can bridge the gap between accuracy and efficiency, enabling scalable deployment of AI-driven attendance systems in resource-constrained educational environments without compromising recognition performance. The achievement of perfect accuracy in real-world conditions validates the practical viability of the proposed approach for smart classroom applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient hybrid architecture for facial recognition on resource-constrained IoT devices in smart classroom applications

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Marwa A. Marzouk

摘要

Deploying deep learning-based facial recognition on resource-constrained IoT devices presents significant challenges due to high computational demands, memory requirements, and power consumption. This paper presents a novel hybrid lightweight architecture for contactless smart attendance systems, specifically optimized for edge computing environments such as IoT-enabled classrooms. Unlike conventional end-to-end deep learning approaches, our system strategically combines deep neural networks with classical machine learning techniques to achieve real-time performance on low-power hardware while maintaining high accuracy. The proposed system employs a five-stage modular pipeline: (1) Motion detection using a low-cost infrared sensor to activate the camera only when movement is detected, reducing power consumption and unnecessary processing. (2) Face detection through an optimized Viola–Jones algorithm employing Haar features, integral images, and a cascade classifier for efficient identification of face regions under constrained resources. (3) Feature extraction using a lightweight FaceNet model trained with triplet loss to generate compact and discriminative 128-dimensional embedding. (4) Face classification using a Support Vector Machine (SVM) that accurately distinguishes identities based on the extracted embedding while minimizing computational load. (5) Automated attendance management with real-time database synchronization and structured logging for administrative verification. Comprehensive experimental evaluation demonstrates superior performance across multiple benchmark datasets: the proposed Viola-Jones + FaceNet + SVM model achieved 98% accuracy on the LFW dataset and 99.8% on the NUAA dataset, outperforming all compared approaches. Most significantly, the system achieved 100% accuracy on our real-world classroom database, demonstrating exceptional practical reliability in authentic deployment scenarios. Implementation on Raspberry Pi 4B achieves an average inference time of 0.68 s per face with only 512 MB memory footprint, representing substantial improvements in computational efficiency compared to end-to-end deep learning alternatives. The system successfully handles real classroom conditions including varying illumination, multiple face angles, and practical deployment constraints. This work demonstrates that carefully designed hybrid architectures can bridge the gap between accuracy and efficiency, enabling scalable deployment of AI-driven attendance systems in resource-constrained educational environments without compromising recognition performance. The achievement of perfect accuracy in real-world conditions validates the practical viability of the proposed approach for smart classroom applications.