Edge-guided slicing and inference (EGSI) with feature aggregation network for quantification of large defective temporary construction materials

Temporary materials, primarily steel tubular sections, are essential for erecting temporary structures on construction sites. Due to frequent reuse and storage conditions, these materials often develop defects such as rust and bends, compromising their safety. While manual quantification of defective materials is time-consuming, especially for batches exceeding 200 materials, computer vision techniques offer a more efficient alternative. However, detecting defects in large batches remains challenging as the materials appear as small objects in images. This paper proposes a feature aggregation network through an ablation study to enhance the detection accuracy of large defective temporary materials and an edge-guided slicing inference method for precise quantification of large batch sizes. The proposed bottleneck layer in the feature aggregation network achieved a mean average precision (mAP) of 81.1%, outperforming the best custom model by 1.2%. Additionally, the edge-guided slicing inference reduced the mean absolute error to 0.1%, compared to 0.6% from the original slicing-aided hyper inference. The developed system was also deployed on a web-based user interface for visualization, improving accessibility and usability. These advancements enhance automated defect detection and quantification of temporary materials, improving efficiency and accuracy in construction management.