Smoking Detection Based on Memristor Convolutional Neural Network

Embedding deep learning models into edge computing systems presents challenges due to hardware limitations, power constraints, and computational complexities. To tackle these, we introduce a memristor-based convolutional neural network (CNN) leveraging a customized lightweight MobileNetV2 architecture tailored for smoking behavior recognition. Our approach employs processing-in-memory (PIM) facilitated by memristor crossbar arrays, significantly reducing computational demand. Tested on a self-constructed dataset, our modified MobileNetV2 achieves an accuracy of 92.25%. Simulation results from MemTorch, considering linear ion drift and realistic Verilog-A RRAM models, yield accuracies of 92.11% and 91.82%, respectively, with minimal performance degradation. This demonstrates the robustness of memristor technology in practical edge computing scenarios.