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).
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