Particle Counting Method for Asbestos Qualitative Analysis Based on Multi-Resolution Analysis

Soichiro Morishita, Takehiro Watanabe, Kuniaki Kawabata, Yoshikazu Arai, Atsushi Yamashita, Hajime Asama, Taketoshi MishimaJournal of the Japan Society for Precision Engineering 2015 · DOI

Abstract

Manual counting of asbestos particles in building materials is a labor-intensive but necessary task for public health safety. This paper proposes an automated particle counting method based on multi-resolution analysis and Support Vector Machines (SVM). By scanning microscope images across a Gaussian pyramid of scales, the system identifies particles of varying sizes and contrasts. Experimental results demonstrate that while individual detection errors (false positives and negatives) exist, they are balanced, resulting in a total particle count that is 102.1% of manual counts. This approach provides a reliable, automated alternative to traditional manual microscopy.

Introduction

Asbestos qualitative analysis is critical for identifying hazardous building materials. Official standards require determining if more than 4 asbestos crystals exist per 3,000 particles in a crushed sample. Currently, this is performed manually by technicians, a process that is exhausting and prone to error. Previous automation attempts using simple image thresholding failed because asbestos particles vary significantly in size, shape, and contrast. This research introduces a multi-resolution scanning method designed to mimic the subjective identification process used by human experts.

Methodology

The core of the proposed method is Multi-Resolution Analysis. Instead of analyzing images at a single scale, the system processes a "Gaussian pyramid" of image scales (x1, x2, x4, and x8).

Multi-resolution analysis of asbestos particles at different scales (x1 to x8).

Feature Extraction and Classification

For every Region of Interest (ROI), the system calculates statistical features, including variance, histograms, gradients, and Higher-order Local Auto-Correlation (HLAC). HLAC is particularly effective as it is shift-invariant, allowing for pattern recognition regardless of a particle's exact position within the ROI. These features are fed into an SVM classifier trained to distinguish the presence of a particle.

Scale Selection

To handle size variability, the system monitors SVM output across scales. A particle is officially counted only when the SVM output reaches a local maximum at a specific resolution, indicating an optimal match between the window size and the particle size. This prevents the over-counting of large particles or the omission of small, faint fibers.

Results

The system was evaluated using 100 full-scale microscope images. The classifier achieved a 97.9% accuracy rate in isolated ROI discrimination. In the full-scale counting test, the automated system identified 3,938 particles compared to the manual count of 3,856—a counting rate of 102.1%.

A significant finding was the symmetry of errors: the False Positive rate (16.9%) and False Negative rate (13.8%) largely canceled each other out. This "error balancing" ensures that the aggregate count remains highly accurate, even if individual detections are imperfect.

Conclusion

The proposed multi-resolution method effectively automates asbestos particle counting with a high degree of parity to manual methods. By treating resolution scale as a dynamic variable and leveraging shift-invariant features, the system overcomes the limitations of traditional thresholding. Future work will address visual artifacts from downsampling and improve performance in densely packed particle clusters.

Summary updated

Abstract

In recent years, the health of many people is damaged by asbestos. Therefore, to detect very small quantity of asbestos that exists in general environment is very important. According to an officially fixed way, existence of asbestos in a certain building material is discriminated according to whether there are more than 4 crystals of asbestos per 3000 particles in samples made with crushing a part of the building material. The particle counting is performed manually, therefore it needs great time and labor. For automation, some studies to count particle with image processing are performed. To ensure consistency with conventional manual method, the automated method should be able to obtain equivalent outcome to manual methods. In this paper, we propose an automatic particle count method based on multiresolution analysis. At first, the classifier which can distinguish existence or nonexistence of a particle at the center of ROI (region of interest) is constructed. Next, the image is scanned with the classifier on multiresolution, and small areas containing a particle are found. To construct the classifier according to the counting result by manual method, we can achieve the result as same as manual one by the proposed method. To confirm validity of the proposed method, we did some experiments to apply the method to the images that particles are identified manually. As the result, we confirmed that 10% of false positive and false negative are contained respectively. Because the proportions are roughly same, the total count is as same as manual counting.
BibTeX
@article{morishita2015particle,
  author = {Soichiro Morishita and Takehiro Watanabe and Kuniaki Kawabata and Yoshikazu Arai and Atsushi Yamashita and Hajime Asama and Taketoshi Mishima},
  title = {{Particle Counting Method for Asbestos Qualitative Analysis Based on Multi-Resolution Analysis}},
  journal = {Journal of the Japan Society for Precision Engineering},
  year = {2015},
  doi = {10.2493/jjspe.81.930},
}
Figure