Feature Extraction of Nanofibers Based on U-Net Multi-classifier

Abstract

 Nanofibrous membrane has great advantages in many fields, of which the micro-structural analysis and optimization are the key to the industrial application. The U-Net multi-classifier based on network structure together with the Jaccard-Lovasz extension loss function was proposed to classify the pixels of the nanofiber SEM image into three categories. A Conditional Random Field (CRF) network was utilized to post-process the segmentation results. Porosities of the filter membranes and the radius of the nanofibers were calculated based on the segmentation results. Experimental results showed that the proposed U-Net multi-classifier could be used to deal with overlapped nanofibers and the corresponding segmentation results could retain important details of the SEM image. The technique is beneficial to the subsequent numerical simulation, which is of great academic and practical significance for the subsequent film performance improvement and application promotion.