Machine vision-based technology for the interface classification of precast concrete components

The interfaces of precast concrete components serve as connections between components. The quality of these interfaces affects not only the bonding strength between the components but also the seismic performance of the entire structure. Therefore, the quality of the interface construction is a crucial factor in ensuring the safety and longevity of engineering projects. To enhance the reliability and objectivity of the inspection results of interface quality, this study introduced machine vision technology and investigated intelligent inspection approaches for interface construction quality, achieving the identification of two hierarchical key targets: construction techniques and roughness grades. First, this study integrated a specialized detection device to collect substantial images in a constrained environment, establishing a stable dataset called the Concrete Techniques and Roughness (CTR) dataset. Second, this study employed a hierarchical Convolutional Neural Network (CNN) structure in which layers make predictions in the descending order of class abstraction based on prior knowledge. This allows the model to classify sequentially according to construction techniques and roughness grades. Additionally, considering the characteristics of interfaces, this study introduced the Fused Strip Pooling and Convolutional (FSPC) attention module as a feature fusion attention mechanism that enhances the ability of the model to focus on the crucial quality features of interfaces. The integration of hierarchical prior knowledge and attention mechanisms significantly improved the classification performance. This approach achieved an accuracy of 100 % for the construction techniques and 96.26 % for the roughness grades on the testing set of the CTR dataset. The results demonstrate that this method provides an effective solution for the application of machine vision in the quality inspection of precast concrete component interfaces.