End-to-End Deployment of Deep Learning Models for Image Classification on Embedded Systems with a Comparative Study of TensorFlow Lite and PyTorch-ONNX Workflows

Naumov, Angel (avtor), Bulić, Patricio (mentor)

This thesis examines the end-to-end deployment of deep learning models for image classification on the STM32H747I-Discovery board, a resource-constrained embedded system. A detailed comparative analysis of two deployment workflows is provided: one using TensorFlow with its native TFLite format and another using PyTorch with ONNX as an intermediate representation. The study documents the process from model training and quantization to on-device inference and evaluates the trade-offs in memory footprint, latency, and development complexity. The results show that the TensorFlow-to-TFLite workflow is a more efficient and streamlined solution for this application. It yields a model with significantly lower RAM usage and faster inference speed. This demonstrates the benefits of a specialized deployment format over a generalized one for memory-critical applications.

• Jezik: Angleški jezik
• Ključne besede: embedded system, artificial intelligence, edge computing, TinyML, STM32
• Vrsta gradiva: Diplomsko delo/naloga
• Tipologija: 2.11 - Diplomsko delo
• Organizacija: FRI - Fakulteta za računalništvo in informatiko
• Leto izida: 2025
• PID: 20.500.12556/RUL-173253
• COBISS.SI-ID: 253218563
• Datum objave v RUL: 15.09.2025

Sekundarni jezik

• Jezik: Slovenski jezik
• Naslov: Celovita implementacija globokih nevronskih mrež za klasifikacijo slik na vgrajenih sistemih z uporabo TensorFlow Lite in PyTorch-ONNX

Izvleček

Ta diplomska naloga obravnava celovit proces implementacije modelov globokega učenja za klasifikacijo slik na mikrokrmilniku z omejenimi viri STM32H747I-Discovery. V nalogi sta primerjana dva ključna pristopa: prvi temelji na ogrodju TensorFlow in formatu TFLite, drugi pa na ogrodju PyTorch in vmesnem formatu ONNX. Z natančno metodologijo in empirično analizo so predstavljene primerjave med porabo pomnilnika, hitrostjo inference in zahtevnostjo razvojnega procesa. Rezultati jasno kažejo, da je pristop s TensorFlow in TFLite učinkovitejši in zmogljivejši za ciljno strojno opremo, kar potrjuje prednosti uporabe specializiranih formatov za vgrajene sisteme.

• Ključne besede: vgrajen sistem, umetna inteligenca, robno računalništvo, TinyML, STM32