Lightweight Dental Image Segmentation with Combined Importance and Redundancy-based Pruning

초록

Deep learning-based tooth segmentation models achieve high accuracy as they grow deeper and more complex; however, their practical application in industrial settings is hindered by high computational and memory demands. To address this, pruning techniques have been employed, but existing methods mainly focus on filter importance and fail to consider redundancy among filters. In this paper, we propose DDGWD, a pruning method that simultaneously incorporates filter importance and redundancy evaluation based on Manhattan distance. Experiments on a tooth segmentation dataset demonstrate that the proposed method effectively reduces computational cost and memory usage compared to conventional importance-based approaches while minimizing performance degradation. This study shows that redundancy-aware pruning can play a crucial role in developing lightweight tooth segmentation models suitable for practical industrial deployment. © 2025 IEEE.

키워드