Comparative Analysis of Lightweight Deep Learning Models for Memory-Constrained Devices