Aplikace neuronových sítí na jednodeskových počítačích pro zpracování signálů v IoT aplikacích

Abstract

This thesis deals with the use of neural networks on single board computers for signal processing in IoT applications, with a focus on image data analysis. The aim is to explore the possibilities of deploying neural networks on hardware-constrained platforms, to optimize their computational complexity and to evaluate their performance in real-world conditions. The theoretical part describes the basic principles of neural networks and their use in IoT. Furthermore, available hardware resources for accelerating computation are presented. The practical part focuses on the detection and recognition of objects in an image using the MobileNetV2 architecture, which is used as the backbone network of the proposed models. These models are trained on two selected datasets with a focus on achieving a balance between accuracy and computational efficiency. Subsequently, the models are deployed on a Raspberry Pi platform with the Google Coral Edge TPU accelerator, which allows speeding up inference computations and reducing processing latency. The experimental results provide a comprehensive overview of the potential applications of neural networks in edge computing.