Deployment of TensorFlow Lite for Real-Time Image Classification on Raspberry Pi: Performance Evaluation with CIFAR-10 and MNIST, and Implications for Power Generation

The present paper investigates the application of TensorFlow Lite to deploy the Convolutional Neural Network on Rasberry Pi for real-time image classification, considering specifically the CIFAR-10 and MNIST datasets. Convolutional Neural Network architectures based on TensorFlow, developed for each dataset under investigation, are hereby employed within a training regime. Thereafter, the trained models were converted already into TensorFlow Lite format to effectively run them on edge devices, such as Raspberry Pi. A camera interface was applied to showcase real-time inference, which proved the viability and effectiveness of implementing machine learning models in contexts that have fewer resources. On the other hand, TensorFlow Lite is designed for smaller, mobile-friendly models with lightweight and effective inference. It is explicitly designed to run on embedded and mobile devices. Both the frameworks support Python and apply the optimization techniques under autotuning, quantization, and model optimization toolkits for optimization targeted to their respective deployment environments. These models further manifest real-time inference through a camera interface and explain their practical implications in the power generation sector. The potential use of lightweight machine learning frameworks is discussed for improving anomaly detection, predictive maintenance, and load forecasting, and in smart grid data analysis in the power sector.