Edge-Deployed Computer Vision for Real-Time Defect Detection

摘要

The key to the development of intelligent manufacturing systems is the realization of precision and low-latency defect detection in real-time. The conventional methods that depend on the centralized cloud computing are prone to network latency, scale and low flexibility of integration, particularly in the industrial environment that has limited connectivity. This paper proposes a computationally efficient computer vision pipeline that can be deployed to the edge. This computer vision pipeline is designed to fill the gap in the demand for high-performance defect detection and the resource limitations of edge systems. This system uses a model of compact deep learning trained with quantization and pruning to achieve low computational complexity without affecting accuracy. The system can easily interface with different kinds of industrial hardware using the modular architecture, such as robotic arms, conveyor systems, and PLCs. We argue for the validity of our methodology using a practical case study that involves detecting surface defects in manufactured metal components. With an inference speed of 30 FPS on the NVIDIA Jetson Nano and a classification accuracy of 97.6%, our answer outperforms classical architectures like ResNet-50 in edge environments. A comparison of the lightweight CNN models, deployment optimizations with TensorRT and system-level performance measures is also provided in the paper. Besides, we solve key problems such as pipeline latency, high resistance to lighting changes, and the reconfigurability of inspection task dynamics. Holistic assessments demonstrate that our deployed solution to the edges is highly efficient and reliable in terms of manufacturing products, but its use can be expanded into such diverse fields as agriculture, healthcare, and unmanageable vehicles. This study opens the door to the use of compact vision systems in such low-capacity spaces, achieving the same level of performance as competing systems