A Wavelet-Based Bilateral Segmentation Study for Nanowires

TL;DR

This paper introduces a new deep learning model for accurately segmenting complex 1D nanowires in images.

Contribution

The novel WaveBiSeNet model with Dual Wavelet Convolution and Flexible Upsampling modules improves nanowire segmentation.

Findings

WaveBiSeNet achieved an mIoU of 77.59% on a peptide nanowire image dataset.

The model outperformed ten other segmentation models in accuracy and F1 score.

Results show WaveBiSeNet effectively handles entangled and blurred nanowire structures.

Abstract

One-dimensional (1D) nanowires represent a critical class of nanomaterials with extensive applications in biosensing, biomedicine, bioelectronics, and energy harvesting. In materials science, accurately extracting their morphological and structural features is essential for effective image segmentation. However, 1D nanowires frequently appear in dispersed or entangled configurations, often with blurred backgrounds and indistinct boundaries, which significantly complicates the segmentation process. Traditional threshold-based methods struggle to segment these structurally complex nanowires with high precision. To address this challenge, we propose a wavelet-based Bilateral Segmentation Network named WaveBiSeNet, to which a Dual Wavelet Convolution Module (DWCM) and a Flexible Upsampling Module (FUM) are introduced to enhance feature representation and improve segmentation accuracy. In this study, we benchmarked WaveBiSeNet against ten segmentation models on a peptide nanowire image dataset. Experimental results demonstrate that WaveBiSeNet achieves, mIoU of 77.59%, an accuracy of 89.95%, an F1 score of 87.22%, and a Kappa coefficient of 74.13%, respectively. Compared to other advanced models, our proposed model achieves better segmentation performance. These findings demonstrate that WaveBiSeNet is an end-to-end deep segmentation network capable of accurately analyzing complex 1D nanowire structures.

Figures