IMAGE CLASSIFICATION ON RASPBERRY PI

Overview

This repository presents a machine learning-based image classification pipeline optimized for edge computing environments like Raspberry Pi.

The system architecture adopts a decoupled computational approach:

• Training and optimize phase is executed on a high-performance host machine
• Trained and optimized model is exported and deployed onto Raspberry Pi

This methodology enables efficient, real-time visual inference with minimal latency without relying on cloud-based processing.

Features

• Edge-based Inference: real-time image classification executed locally on the Raspberry Pi hardware.
• Host training: Utilization of host computing resources for robust model training, hyperparameter tuning, and weight optimization.
• Model optimization : integration of pruning techniques to reduce model footprint, benchmarking to ensure an optimal balance between accuracy and computational efficiency.

Repository Structure

• data_prep: create dataset and pre-processing scripts for training step
• train: training and package models script
• inference: deploy, inference and benchmark
• eval: performance comparison between baseline model and pruned model

System Specifications

The system is developed and evaluated on the following hardware and software stack:

• Edge Hardware: Raspberry Pi 4 Model B - 8GiB RAM with Logitech Webcam C505E
• Software Environment: Raspberry Pi OS, Python 3.11, TensorFlow 2.x, OpenCV, and TensorFlow Lite for edge inference.