Autonomous microscopic bunch inspection using region-based deep learning for evaluating graphite powder dispersion. (10th June 2018)

• Record Type: Journal Article
• Authors: Liu, Hai; Sha, Aimin; Tong, Zheng; Gao, Jie
• Is Part Of: Construction & building materials. Volume 173(2018)
• Page Start: 525
• Page End: 539
• Publication Date: 2018-06-10
• Subjects: Graphite powder bunch -- Dispersion of graphite powder -- The Faster R-CNN -- Optical microscope (OP) -- Region proposal networks; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2018.04.050
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: Project presents an application of convolutional neural networks (CNN) in evaluating graphite powder dispersion. Different Faster R-CNNs are established by the processes of structure design, training and testing. The optimal well-trained Faster R-CNN is able to locate graphite powder bunch with acceptable precision and high efficiency. The method based on the Faster R-CNN had the capacity of quasi real-time autonomous dispersion evaluation in GPU mode. Abstract: The ice-snow melting performance of ice-snow pavement is significantly influenced by the dispersion of graphite powder, particularly through the distribution of graphite powder bunches. In recent years, optical microscope (OP) images have been utilized to detect graphite powder bunches and evaluate their dispersion. However, because graphite powder bunches and other objects in OP images often have various shapes, and conventional manually processed images of tasks have the disadvantage of low efficiency, it is a challenge to detect graphite powder bunches and evaluate their dispersion using OP images. Therefore, this paper presents a novel application of a Faster Region Convolutional Neural Network (Faster R-CNN) using OP images and video sequences for the autonomous detection of graphite powder bunches and an evaluation of their dispersion. The research procedure is as follows: (a) generate a database for the Faster R-CNN, (b) design 30 Faster R-CNNs to find the optimal one, and (c) conduct an analysis of the training and testing results, along with new image testing, comparative studies, and video testing. The results show that a Faster R-CNN with nine anchors and a ratio of 0.3, 1.0, and 1.6, and with the sizes of 32, 128, and 192, has an average precision of the bunches and a dispersion uniformity of 91.2% and 84.0%, respectively. Its mean average precision is 87.5%. The Faster R-CNN is considered optimal in this research. The test time required to evaluate an image with a pixel resolution of 1024 × 1024 pixel in GPU mode is approximately 0.04 s, which means the method based on a Faster R-CNN has the capacity of a quasi-real-time autonomous dispersion evaluation in GPU mode to replace a human-assisted microscopic dispersion evaluation in OP images. The results also provide the possibility for a quasi-real-time evaluation using OP video sequences. Compared with a Fast R-CNN, a Faster R-CNN provides more reasonable bounding boxes for bunches and reliable results in terms of the dispersion uniformity.