Quality Assessment of Composite Manufacturing Using Automated Visual Inspection System with SEM Images

The morphology of fibers (e.g., spatial uniformity, orientation, and length) plays a decisive role in determining the material properties or fabrication quality of fiber-reinforced nanocomposites. Hence, determining the morphology becomes a very critical issue in the field of nanocomposite quality control. The conventional way of quality inspection is to take the scanning electron microscopic (SEM) images of the cross-section of composite material and do the visual checking of these SEM images to evaluate the nanofiber alignment and length distribution. But this type of inspection is often subjective, inaccurate and time consuming. Moreover, the extremely small size of nanofibers makes the quality control evaluation process very tedious. These challenges are not well addressed in the existing literature. This dissertation responds to this gap in knowledge and develops several elaborate methodologies to automate the tasks of quality assessment of short fibers reinforced composite manufacturing. This research consists of two broad steps: fiber segmentation and quantitative analysis of segmented fibers. In the first step, we develop five different methods, namely, the opening method, simple Hough Transform (HT), partitioning HT, gradient-based HT, and break-merge method to automatically extract the short straight fibers from SEM images. Later, in the second step, the extracted fibers are quantitatively analyzed to facilitate the morphological analysis. Extraction of filler-morphology greatly depends on accurate segmentation of fillers (fibers and particles). Convolution Neural Networks (CNNs) have been shown to be very effective at object recognition in digital images. Realizing the potentials of CNNs, in this dissertation, we also propose an automatic filler detection system in SEM images, utilizing a Mask Region-based CNN architecture. The proposed system can simultaneously classify, detect, and segment fillers in SEM images, making it suitable for morphology analysis of fillers and automatic quality inspection. We also propose a novel SEM image simulation procedure to overcome the data scarcity for training a deep CNN architecture. The proposed filler detection system is trained on the simulated images. It is shown that the trained network can detect and segment fillers with higher accuracy even in the overlapping and obscure situations. The performance and robustness of the proposed system are evaluated using both simulated and real microscopic images