CrackClassification4: Classifying Four Types of Cracks in Structural Surfaces Using CNNs

This study addresses the critical task of crack identification in structures like walls and pavements, essential for preventing deterioration and ensuring safety. Traditional inspection methods have proven slow and inaccurate, leading to the need for a more efficient machine learning solution. Despite advances in crack detection, specific crack classification has remained underexplored. To bridge this gap, the research introduces a classification system for four types of cracks: hairline, diagonal, alligator, and longitudinal. An extensive dataset CrackClassification4 was created, and six pre-trained CNN models—VGG16, VGG19, InceptionV3, MobileNetV2, and ResNet50—were trained using transfer learning. Additionally, a custom CNN model was developed, featuring improved feature extraction and a deeper architecture. Evaluation of the models showed the custom CNN outperformed the rest, achieving an accuracy of 0.98, precision of 0.98, recall of 0.98, and an F1-score of 0.98. MobileNetV2 followed with 0.94 accuracy and a 0.95 F1-score, making it suitable for real-time applications with limited computational resources. The findings confirm that CNN-based models, especially the custom CNN, can significantly enhance the accuracy of infrastructure monitoring and crack detection.