Optimized for the Edge: A Lightweight AI Model for Industrial Product Classification

Abstract

This paper presents EdgeSorter: an end-to-end embedded AI system for real-time industrial product classification and sorting. Deployed on the ESP32-CAM microcontroller, our optimized MobileNetV1 architecture achieves 95% classification accuracy for three fruit types while operating within extreme resource constraints (137.7 KB RAM, 82.2 KB Flash). Through systematic quantization and hardware-aware optimization, we demonstrate real-time inference at 45 ms latency, enabling direct mechatronic control of sorting mechanisms. Extensive experiments validate the system's operational efficacy, with comparative analysis showing 15% improvement over traditional computer vision approaches. This work provides both a technical blueprint and performance benchmarks for practical edge AI deployment in Industry 4.0 applications, bridging the critical gap between algorithmic potential and embedded realization.