FIBER SEGMENTATION IN COMPOSITE MATERIALS USING MARKED POINT PROCESSES
Résumé
This paper presents a new method dedicated to unsupervised 2D segmentation of fibers in a section of composite carbon-fiber materials. The framework relies on a marked point process algorithm. We shall create random elliptical objects to fit the fiber distribution in the image. The interaction rules between the objects complete the model. Using a Markov Chain Monte Carlo (MCMC) method, the algorithm converges to a configuration which is close to the fiber distribution in the images. At each step, the configuration is evaluated considering its proximity to the target distribution. In order to achieve this task, we propose a mixed data model using both grey level values and gradients to evaluate the likelihood of the current configuration. This mixed model overcomes the problems of luminance variation, contour discontinuities and high noise level. Finally, the results on the composite material sections illustrate the efficiency of the segmentation and suggest that the marked point processes can be a promising tool for fiber detection.