A graph attention reasoning model for prefabricated component detection

Abstract

Accurately checking the position and presence of internal components before casting prefabricated elements is critical to ensuring product quality. However, traditional manual visual inspection is often inefficient and inaccurate. While deep learning has been widely applied to quality inspection of prefabricated components, most studies focus on surface defects and cracks, with less emphasis on the internal structural complexities of these components. Prefabricated composite panels, due to their complex structure—including small embedded parts and large‐scale reinforcing rib—require high‐precision, multiscale feature recognition. This study developed an instance segmentation model: a graph attention reasoning model (GARM) for prefabricated component detection, for the quality inspection of prefabricated concrete composite panels. First, a dataset of prefabricated concrete composite components was constructed to address the shortage of existing data and provide sufficient samples for training the segmentation network. Subsequently, after training on a self‐built dataset of prefabricated concrete composite panels, ablation experiments and comparative tests were conducted. The GARM segmentation model demonstrated superior performance in terms of detection speed and model lightweighting. Its accuracy surpassed other models, with a mean average precision (mAP₅₀) of 88.7%. This study confirms the efficacy and reliability of the GARM instance segmentation model in detecting prefabricated concrete composite panels.