Application of deep learning-based image recognition technology to asphalt–aggregate mixtures: Methodology. (23rd August 2021)

• Record Type: Journal Article
• Authors: Dan, Han-Cheng; Bai, Ge-Wen; Zhu, Zhi-Heng
• Is Part Of: Construction & building materials. Volume 297(2021)
• Publication Date: 2021-08-23
• Subjects: Deep learning -- Image recognition -- Asphalt mixture -- Aggregate shape -- U-NET++; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2021.123770
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: Image reconstruction compositing is used to produce image of asphalt–aggregate mixture. Deep learning-based method is used to collect information of asphalt mixture aggregate. Training and recognition is improved using Gaussian blur and secondary labeling methods. Identifying and analyzing segregation and weakness in the asphalt mixture is proposed. Image data can be used to quickly derive index parameters for the asphalt mixture. A lightweight and intelligent analysis method is proposed for asphalt–aggregate mixtures. Abstract: An important factor that affects the performance and durability of asphalt–aggregate mixtures is the size distribution of the aggregate particles. However, there are no accurate and quick methods of obtaining this information currently available. A potential solution to this problem is to use computer-aided image recognition technology, a technique that is being used increasingly more often in the field of pavement engineering. In this paper, asphalt core samples drilled from pavements are used as test objects. The information collected from the side surfaces of these cores is analyzed using neural network technology to help improve the speed and accuracy of the quantitative analysis of the asphalt mixtures. First, a camera is used to make a series of photographic images of the side surface and image reconstruction compositing is adopted to produce an image of the asphalt–aggregate mixtures. This image can be processed using a neural network (U-NET++) to identify the aggregate in the image. The training and recognition effect is improved by using Gaussian blur and secondary labeling methods. The grayscale images can then be binarized, filled, and the aggregate particles are segmented in order to identify holes and areas with adhesion problems. The geometric forms of the particles that appear on the side surfaces of the asphalt mixtures can be accurately obtained. Important geometric parameters (e.g. aspect ratio and roundness) can thus be derived from the image. A method of identifying and analyzing segregation and weakness in the asphalt mixture is proposed. The results show that our composite images are more accurate than traditionally-synthesized panoramic images. Also, the trained U-NET++ is ideal for recognizing and segmenting the aggregate image. The data obtained from the image can thus be used to quickly derive index parameters for the asphalt mixture. Compared with traditional test methods, this method only requires the use of mobile phones and personal computers. It is thus a lightweight and intelligent analysis method for acquiring information about asphalt–aggregate mixtures.