1A1-D06 Asbestos Detection Method with Frequency Analysis for Microscope Images

Hikaru Kumagai, Soichiro Morishita, Kuniaki Kawabata, Hajime Asama, Taketoshi MishimaThe Proceedings of JSME Annual Conference on Robotics and Mechatronics (Robomec) 2009 · DOI

Abstract

Increased health damage caused by asbestos has become a significant concern in recent years. Mitigating these hazards requires the accurate identification of asbestos in building materials. Current official protocols involve manual counting of particles in microscope images of pulverized samples, a task that is both time-consuming and labor-intensive. While automated image processing systems are being developed to support this process, they often struggle to detect "low-brightness" asbestos crystals, which are hazardous but difficult to distinguish from the background. This study proposes an asbestos detection method that focuses on the frequency distribution of images, enabling the identification of these stealthy fibers based on their geometric orientation rather than light intensity.

The Challenge of Manual Asbestos Detection

Identifying asbestos in building materials is a critical but grueling task. Official safety protocols require technicians to manually count up to 3000 particles across multiple samples under a microscope. If four or more asbestos crystals are identified, the material is deemed hazardous. This manual process is slow, exhausting, and prone to human error due to the high volume of particles that must be scrutinized.

Manual inspection of samples is a time-consuming and labor-intensive task for technicians.

Existing automated systems typically rely on brightness thresholds. However, asbestos crystals are often needle-like and can appear very dim or have low contrast against the background. Standard software often ignores these low-brightness particles to avoid noise, potentially leading to a "false safe" reading for dangerous materials. The researchers sought a method to detect these fibers based on their distinct shape and orientation.

Methodology: Frequency-Domain Analysis

The proposed method shifts the focus from pixel intensity to geometric frequency through a multi-step pipeline:

  1. Grayscale Conversion: Instead of standard weighted conversion, the system uses a simple average of RGB channels $(R+G+B)/3$. This ensures that low-brightness blue or yellow crystals—common in dispersion staining microscopy—are not filtered out prematurely.
  2. Frequency Analysis (2D FFT): The image is divided into $32 \times 32$ pixel blocks, each undergoing a 2D Fast Fourier Transform (FFT). Because asbestos crystals are elongated, they create a distinct, directional signature in the frequency domain.
  3. Eigenvalue Decomposition: By calculating a covariance matrix of the frequency distribution and finding the primary eigenvector, the system mathematically determines the specific orientation of the crystal within each block.
  4. Directional Edge Enhancement: Using the calculated angle, a $5 \times 5$ Sobel filter is applied, oriented specifically to enhance edges at that angle. This "tunes" the detector to the specific fiber identified.
  5. Binarization: The enhanced image is converted to black and white for final counting.
A typical microscope image used in the study, containing low-contrast and low-brightness particles.

Experimental Results

The study utilized a dataset of 21 microscope images ($640 \times 480$ pixels) provided by RIKEN, containing 889 particles and 40 confirmed asbestos crystals. The proposed method achieved the following performance metrics:

  • Asbestos Detection Rate: 92.5% (37 out of 40 crystals detected).
  • Particle Counting Rate: 107.6% (957 particles detected vs. 889 actual).
  • False Detection Rate: 18.6%.

The high detection rate for asbestos confirms the effectiveness of using frequency analysis to find low-brightness fibers. However, the system tended to over-count, occasionally mistaking air bubbles or image noise for hazardous particles.

Conclusion and Future Work

By prioritizing geometry over brightness, this method successfully identifies hazardous asbestos fibers that are often missed by conventional automated systems. The core innovation lies in using FFT to identify the "streak" created by needle-like objects regardless of their light intensity. Future research will focus on reducing the false positive rate by improving performance in crowded image regions and optimizing resolution to further distinguish asbestos from non-hazardous noise.

Summary updated

Abstract

Increased health damage caused by asbestos has become a problem recently. Removal of asbestos from building materials or rendering it harmless are common means of alleviating asbestos hazards. In the process, a judgment of whether asbestos are included in building materials is required. Complying with an official method, particles and asbestos in microscope images of samples made from pulverized building materials should be counted to judge whether the samples contain asbestos. This task is performed manually and requires enormous amount of time and burden. Asbestos containing judgment support system with image processing is being developed to improve its efficiency. However, low brightness particles must not be detected because of judging not harm but harmless, detecting low brightness asbestos is difficult. This study is intended to establish a method is also able to detect low brightness asbestos, we propose an asbestos detection method focusing on frequency distribution of the image.
BibTeX
@inproceedings{kumagai2009d,
  author = {Hikaru Kumagai and Soichiro Morishita and Kuniaki Kawabata and Hajime Asama and Taketoshi Mishima},
  title = {{1A1-D06 Asbestos Detection Method with Frequency Analysis for Microscope Images}},
  booktitle = {The Proceedings of JSME Annual Conference on Robotics and Mechatronics (Robomec)},
  year = {2009},
  doi = {10.1299/jsmermd.2009.\_1a1-d06\_1},
}
Figure