Post-Training Quantization for U-Net Based Segmentation: Model Compression and Performance Evaluation

As edge devices and CPU-based systems become increasingly important, efficient inference on resource-constrained hardware is a critical challenge. U-Net, a widely used segmentation model in medical imaging, infrastructure inspection, and satellite image analysis, is often too computationally heavy for such environments. This study explores the application of Post-Training Quantization (PTQ) to reduce the complexity of U-Net without requiring retraining. We train the U-Net model in full-precision (FP32) and then apply several PTQ scenarios, including full-model quantization and selective layer-wise quantization. We evaluate the trade-off between model compression and segmentation performance across different quantization levels. Experimental results demonstrate that PTQ can significantly reduce model size and inference time, with minimal accuracy loss in specific configurations. Our findings highlight the practical potential of deploying U-Net-based models on edge devices using quantization techniques.