Deep learning assisted particle size ranking and estimation from SEM images without explicit segmentation.

Area of Science:

• Materials Science
• Nanotechnology
• Image Analysis

Background:

• Accurate particle size and morphology determination from Scanning Electron Microscopy (SEM) images is crucial in nanomaterial research.
• Conventional methods often require laborious manual segmentation or parameter tuning, especially for complex particle structures.

Purpose of the Study:

• To develop a novel, segmentation-free framework for ranking and estimating particle size ratios from SEM images.
• To leverage a similarity-driven deep learning approach using synthetic data for training.

Main Methods:

• A Convolutional Neural Network (CNN) was trained exclusively on synthetically generated SEM-like images.
• The network learns a morphology-aware embedding space encoding implicit particle size information.
• Particle size ratios are estimated through embedding-based retrieval and similarity ranking.

Main Results:

• The framework achieves robust ranking performance and visual consistency on real SEM datasets.
• It successfully bypasses the need for pixel-level ground truth masks and manual annotation.
• Demonstrates operator-independent, rapid screening of materials.

Conclusions:

• The proposed segmentation-free framework offers a practical alternative to traditional metrological segmentation for SEM image analysis.
• This approach enables efficient and accurate relative size assessment of nanoparticles.
• Facilitates rapid material screening in nanomaterial research.