Performance Contribution and Quantization Sensitivity-based Quantization Strategy for Hybrid Dental Image Segmentation Network

초록

Deep learning-based dental image segmentation algorithms have received considerable attention due to their high performance. In particular, hybrid network-based approaches that combine convolutional neural networks (CNNs) with vision transformers demonstrate superior performance compared to single-network models. Although quantization techniques have been explored to enable the deployment of these algorithms in resource-constrained hardware environments, research on quantization for hybrid networks remains limited relative to that for single-network architectures. In this paper, we investigate strategies for efficiently quantizing hybrid networks by analyzing each block's contribution to overall network performance and its sensitivity to quantization. © 2025 IEEE.

키워드