Using Computer Vision to Recognize Construction Material: A Trustworthy Dataset Perspective

Ying Sun,Zhaolin Gu

DOI: https://doi.org/10.1016/j.resconrec.2022.106362

IF: 13.716

2022-01-01

Resources Conservation and Recycling

Abstract:Data quality is vital in machine learning models, and data scientists spend a substantial amount of time on data quality certificates before model training. With the objective of building trustworthy construction material datasets in this study, we first built construction material datasets, namely those for concrete, brick, metal, wood, and stone data, by collecting images of different construction materials. In particular, these datasets include the main construction material (i.e., concrete, brick, metal, wood, and stone), background, and uncertainty cate-gories. Subsequently, we propose a novel cross-review framework that applies the entire corresponding datasets as the input (e.g., brick datasets) and returns clean datasets. Subsequently, several state-of-the-art deep con-volutional neural networks such as VGG16, GoogleNet, and ResNet were selected to verify the quality and effectiveness of the construction material datasets under two pipelines: friendly mobile devices and unfriendly architectures. The experimental results show that the proposed construction waste material datasets have satisfactory quality and can be effectively recognized by these state-of-the-art deep neural network models. Additionally, the results indicate the necessity to clean the data before training a deep-learning model. Further research should be combining a novel deep construction material recognition model with the cross-review framework to further improve the recognition performance.