Influence of data source and volume on CNN applications in construction

Abstract

Convolutional Neural Networks (CNNs) are widely used in construction. However, the impact of data characteristics on their performance remains underexplored. This review aims to summarize previous papers from the perspectives of data source and volume and to understand their relationship with accuracy. A literature review of 162 papers revealed that 75% of the papers utilized data from a single source, often resulting in higher performance than multiple and public sources. The mean sample numbers of models with and without pre-training were 9,307 and 554,305, respectively. The review results indicated that the relationship between the number of samples and accuracy was moderately positive, and pre-training may allow for performance improvement even with fewer samples. This review highlights efforts toward improving publicly available data and pre-trained models in the construction community and using diverse data sources for validation to ensure the generalization of CNN algorithms for practical applications.