Improved U-Net-Based Pore Segmentation Method for Nano-Ag Coupling Layer
Yiqing Gu, Mingyuan Wang, Jiuhong Jia, Shan‐Tung Tu · IEEE Sensors Journal · 2025
Nano-Ag coupling layer has been successfully applied in ultrasonic transducers, where the morphology of internal pores significantly affects the coupling performance. In this work, a nano-Ag coupling layer pore segmentation method based on improved U-Net is proposed. The proposed method enhances the traditional U-Net architecture by incorporating atrous spatial pyramid pooling (ASPP), squeeze-and-excitation (SE) blocks, attention mechanisms, and deep supervision strategies to augment the model’s feature extraction and representation capabilities. Experiments conducted on a dataset of scanning electron microscope (SEM) images of nano-Ag coupling layers demonstrate that the proposed model outperforms classic segmentation models, such as U-Net, achieving a mean intersection over union (Mean IoU) of 0.663 (compared to 0.580), a mean dice coefficient of 0.759 (compared to 0.700), and other superior metrics, including a mean precision of 0.793, a mean pixel accuracy of 0.988, and a lower mean Hausdorff distance of 4.857 (compared to 6.918). Ablation studies further verify the contributions of each improvement module to the model’s enhanced performance. The methodology proposed in this work provides an effective solution for the precise segmentation of pores in nano-Ag coupling layers, supporting quantitative characterization and prediction of the coupling performance in ultrasonic transducers.