Semi-supervised Large-scale Fiber Detection in Material Images with Synthetic Data

TL;DR

This paper introduces a semi-supervised deep learning approach for large-scale elliptical fiber detection in microscopic images, leveraging synthetic data and domain adaptation to improve robustness and reduce annotation effort.

Contribution

It presents a novel semi-supervised method with domain adaptation, RoI-ellipse learning, and symmetry-based ranking for fiber detection in degraded material images.

Findings

Effective in detecting fibers in real microscopic images

Reduces need for extensive manual annotations

Robust to various image degradations

Abstract

Accurate detection of large-scale, elliptical-shape fibers, including their parameters of center, orientation and major/minor axes, on the 2D cross-sectioned image slices is very important for characterizing the underlying cylinder 3D structures in microscopic material images. Detecting fibers in a degraded image poses a challenge to both current fiber detection and ellipse detection methods. This paper proposes a new semi-supervised deep learning method for large-scale elliptical fiber detection with synthetic data, which frees people from heavy data annotations and is robust to various kinds of image degradations. A domain adaptation strategy is utilized to reduce the domain distribution discrepancy between the synthetic data and the real data, and a new Region of Interest (RoI)-ellipse learning and a novel RoI ranking with the symmetry constraint are embedded in the proposed method. Experiments on real microscopic material images demonstrate the effectiveness of the proposed approach in large-scale fiber detection.