Attention-augmented deep architectures for high-precision cement surface crack detection
DOI:
https://doi.org/10.18488/76.v13i4.5371Keywords:
Automated visual inspection, Cement surface crack detection, Convolutional block attention module, Deep learning, InceptionV3, Structural health monitoring.Abstract
This study aims to improve the reliability and efficiency of automated cement surface crack detection by evaluating attention-augmented deep learning architectures for structural health monitoring. Accurate crack identification is essential for maintaining infrastructure safety, reducing manual inspection errors, and supporting timely maintenance decisions. Traditional visual inspection methods are time-consuming, subjective, and prone to human fatigue, whereas conventional convolutional neural networks misclassify complicated surface textures, stains, and variations in illumination as structural cracks. To overcome these limitations, we present and compare three deep learning architectures under identical experimental conditions. Performance was achieved with a conventional CNN based on InceptionV3, with additional performance improvements obtained by incorporating a CBAM module into InceptionV3 and into VGG16. The convolutional block attention module applies spatial and channel attention in series to improve representations and direct the network to regions of interest while suppressing background noise. All models were trained and tested on a balanced set of 4000 high-resolution images of cement surfaces with both cracked and non-cracked samples. Results confirmed that the CBAM + InceptionV3 architecture provided the highest overall performance, with 99.17% accuracy and 99.33% sensitivity, and a reduction in false positives relative to the conventional CNN. In addition, the CBAM + VGG16 architecture achieved comparable accuracy with substantially lower computational requirements and was therefore well-suited for applications with limited resources. The practical implication of this study is that attention-based feature refinement can enhance decision reliability, reduce inspection uncertainty, and support the development of efficient automated systems for civil infrastructure monitoring.
