نوع مقاله : مقاله پژوهشی
عنوان مقاله English
نویسندگان English
Structural health monitoring and timely maintenance are important for building safety, and crack patterns can provide useful information about the probable origin of damage. This study presents a computer-vision and deep-learning framework for automated classification of slab and wall cracks to support preliminary building-damage assessment. The dataset contains 1,292 images representing ten target crack types, including five slab-crack types and five wall-crack types, together with one no-crack reference class for each group. The class distribution is not fully balanced, ranging from 71 to 145 images per slab class and from 72 to 178 images per wall class. Preprocessing includes Gaussian filtering, Canny edge detection, weighted image fusion, and image resizing, and an EfficientNetB0 backbone is trained using transfer learning. Model evaluation in this study is based on Accuracy and Loss trends for the training and validation data. The final validation accuracies are 98.18% for the slab model and 96.62% for the wall model. The main contribution is the multiclass classification of crack patterns in two building components and the linkage of model outputs to engineering information, including probable crack origin, structural/non-structural classification, and preliminary risk guidance for selected structural cracks through a graphical user interface. The proposed system is intended as a support tool for preliminary screening rather than a replacement for professional structural inspection.
کلیدواژهها English
