Resource-Efficient Deployment of Deep Learning Models for Semiconductor Defect Classification on Embedded Devices

Powerful deep learning (DL) models for semiconductor defect classification are computationally infeasible for deployment on low-power edge devices. This research addresses this gap by analyzing the effects of pruning and post-training quantization on model accuracy, latency, and power consumption on a Raspberry Pi. Through multi-objective analysis, we found that 8-bit integer quantization proved highly effective, reducing inference latency by 41% (from 3.06 s to 1.79 s) and average power consumption by 50% (from 15.9 W to 7.6 W), while maintaining or slightly improving accuracy to 94.1%. The optimal deployment configuration, which combines 8-bit quantization with moderate pruning ($40 \%-60 {\%}$sparsity), reduces energy per inference by approx. 3.5 times and decreases model size from 90 MB to 25 MB. These findings validate that production-grade accuracy can be maintained, enabling practical, on-device defect detection under stringent edge computing budgets.