Faqih Ma'arif, Han Ay Lie, Slamet Widodo, Zhengguo Gao, Fardiansyah Nur Aziz, Maris Setyo Nugroho
Crack detection in concrete structures is crucial for maintaining safety and preventing structural damage. Traditional manual inspection methods have limitations in terms of efficiency, objectivity, and coverage on large-scale infrastructure. This study proposes an automated crack segmentation approach using a fully convolutional network with a modified architecture from the Visual Geometry Group. This method applies hierarchical feature learning over multiple deep-trained side outputs and addresses class imbalance with a balanced cross-entropy-based loss function. The training dataset consists of 300 real images representing road surfaces, walls, and concrete structures with real cracks, and 300 binary images as ground truth. Evaluations were conducted on various concrete block shapes, including square, T-shaped, and cylindrical forms. The third side output demonstrated the best performance with an F1-score of 83.3%, a precision of 77.0%, and a recall of 90.8%. The linear fusion strategy can effectively integrate multi-level features, resulting in an average Intersection over Union of 80.4%. The proposed model shows significant improvement over previous methods and can recognize crack patterns across various scales and structural shapes. These results confirm the potential of the proposed approach as a solid basis for automated infrastructure inspection systems. © (2026), (Dr D. Pylarinos). All rights reserved.
Department of Civil Engineering, Universitas Negeri Yogyakarta, Sleman, Indonesia; Department of Civil Engineering, Diponegoro University, Semarang, Indonesia; School of Transportation Science and Engineering, Beihang University, Beijing, China; Department of Electronics and Informatics Engineering, Universitas Negeri Yogyakarta, Sleman, Indonesia
