Performance Evaluation of Image Classification Models on Resource-Constrained STM32 Microcontrollers

Abstract

Deploying deep learning on microcontrollers offers real-time intelligence at the edge, but tight memory and compute budgets complicate design choices. This study evaluates image classification on the STM32H747I-DISCO using a compact convolutional neural network trained on five board classes (Arduino Uno, Node MCU, ESP8266-01, Micro: bit V2.0, ESP32-CAM). A small, augmented dataset (50–100 images per class) was used with standard transformations; models were quantised to int8 and deployed via STM32CubeIDE and the STM32-AI CLI. The analysis examines how input resolution (1080p vs 480p) interacts with accuracy, memory footprint, latency, and power. Four classes achieve ≥95% accuracy across both resolutions, while ESP8266-01 improves from 65.7% (1080p) to 92.3% (480p), suggesting that downsampling can suppress distracting fine-grained artefacts. Activation-buffer tuning and post-training quantisation reduce RAM from ~761 kB to ~610 kB and Flash from ~1.42 MB to ~1.20 MB without accuracy loss; 480p further lowers latency by up to 35% and power by ~20%. The findings provide a resolution-aware benchmark and practical guidance for balancing fidelity and efficiency on STM32-class MCUs, and they motivate future work with larger benchmarks, cross-platform comparisons, and pruning/distillation pipelines