Abstract: The orientation of fibers has a significant impact on the macroscopic properties of short fiber reinforced composite materials. The fiber orientation index is extracted from scanning electron microscope (SEM) images by machine vision methods. To build the training/testing data sets, a fiber orientation distribution is derived based on the orthogonal elliptic closed approximation, and then many simulated SEM images are generated by using the acceptance-rejection and random sequential adsorption algorithms. Based on these simulated images, a BP neural network based on gray-level co-occurrence matrix (GLCM-BP) was proposed to predict the fiber orientation index, and the results were compared with commonly used methods, including morphological segmentation, structure tensor relationship, and convolutional neural network (CNN) algorithms. The results showed that the GLCM-BP model could effectively predict the fiber orientation with a fitting correlation of 0.99 and a mean square error of approximately 0.01, meeting engineering requirements. In comparison, the structure tensor formula systematically underestimates the orientation in planar distributions; morphological and GLCM-BP methods perform better for low fiber volume fractions; GLCM-BP and CNN methods perform better for high fiber volume fractions. The proposed GLCM-BP method also shows capability to resist image noise.

Key words: short fiber reinforced composites, orientation index, artificial neural network, morphology, computer vision