온디바이스 비전 검사를 위한 경량 AI 프레임워크 성능 분석
Performance Analysis of Lightweight AI Frameworks for On-Device Vision Inspection
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper analyzes the performance of lightweight AI frameworks in edge computing environments, addressing the growing need for on-device vision inspection due to its benefits in real-time processing, data privacy, and reduced network latency. The study compares the performance of almond defect detection models based on MobileNetV2 and MobileNetV3 Small, implemented using lightweight PyTorch and TensorFlow Lite on a Raspberry Pi 4 Model B. The main evaluation metrics include model file size, memory usage, inference time, accuracy, precision, recall, and F1-score. Results indicate that TensorFlow Lite with post-training quantization generally achieves the smallest model size and fastest inference speed. However, lightweight PyTorch models demonstrate superior accuracy. The study demonstrates the potential for industrial applications of edge AI in vision inspection tasks and emphasizes the need for further research on various hardware platforms, optimization techniques, and lightweight model architectures to enhance the efficiency and effectiveness of edge AI solutions.
Abstract
This paper analyzes the performance of lightweight AI frameworks in edge computing environments, addressing the growing need for on-device vision inspection due to its benefits in real-time processing, data privacy, and reduced network latency. The study compares the performance of almond defect detection models based on MobileNetV2 and MobileNetV3 Small, implemented using lightweight PyTorch and TensorFlow Lite on a Raspberry Pi 4 Model B. The main evaluation metrics include model file size, memory usage, inference time, accuracy, precision, recall, and F1-score. Results indicate that TensorFlow Lite with post-training quantization generally achieves the smallest model size and fastest inference speed. However, lightweight PyTorch models demonstrate superior accuracy. The study demonstrates the potential for industrial applications of edge AI in vision inspection tasks and emphasizes the need for further research on various hardware platforms, optimization techniques, and lightweight model architectures to enhance the efficiency and effectiveness of edge AI solutions.
참고문헌 (17)
- M. Satyanarayanan, "The Emergence of Edge Computing", Computer, 50(1), pp.30~39 (2017)
- W. Shi, "Edge Computing : Vision and Challenges", IEEE Internet of Things Journal, 3(5), pp.637~646 (2016)
- Z. Zhou, "Edge Intelligence : Paving the Last Mile of Artificial Intelligence with Edge Computing", Proceedings of the IEEE, 107(8), pp.1738~1762 (2019)
- L. Ye, "The Challenges and Emerging Technologies for Low-Power Artificial Intelligence IoT Systems", IEEE Transactions on Circuits and Systems I : Regular Papers, 68(12), pp.4821~4834 (2021)
- C. Profentzas, "Performance of Deep Neural Networks on Low-power IoT Devices", Proceedings of the Workshop on Benchmarking Cyber-Physical Systems and Internet of Things, pp.32~37 (2021)
- F. MohiEldeen Alabbasya, "Compressing Medical Deep Neural Network Models for Edge Devices Using Knowledge Distillation", Journal of King Saud University-Computer and Information Sciences, 35(7), pp.101616~ (2023)
- Aparna, "An Overview of Industrial Vision Systems", International Journal of Advance Research and Innovative Ideas in Education, 3(4), pp.201~207 (2017)
- Y. Zhao, "Automatic Recognition of Surface Defects of Hot Rolled Strip Steel Based on Deep Parallel Attention Convolution Neural Network", Materials Letters, 353, pp.135313~ (2023)
- E. Li, "Edge AI : On-Demand Accelerating Deep Neural Network Inference via Edge Computing", IEEE Transactions on Wireless Communications, 19(1), pp.447~457 (2020)
- A. Howard, "Searching for MobileNetv3", Proceedings of the IEEE, 2019, 1314~1324
- M. Yurdakul, "Flower Pollination Algorithm-Optimized Deep CNN Features for Almond(Prunus dulcis) Classification", Proceedings of International Conference on Emerging Systems and Intelligent Computing, pp.433~438 (2024)
- W. Glegoła, "MobileNet family tailored for Raspberry Pi", Procedia Computer Science, 192, pp.2249~2258 (2021)
- J. Ryu, "Experimental Analysis of The MobileNetV2 Lightweight Method for Leveraging Deep Learning Model in Mobile Devices", Proceedings of the Conference of The Korean Institute of Information Scientists and Engineers, pp.633~635 (2020)
- "Image Classification with MobilenetV2, Arm NN, and TensorFlow Lite Delegate pre-built binaries, Version 21.11Tutorial"[URL]