Crack detection in structural images using a hybrid Swin Transformer and enhanced features representation block

Introduction: This paper presents a crack detection framework employing a hybrid model that integrates the Swin Transformer with an Enhanced Features Representation Block (EFRB) to precisely detect cracks in images. Methods: The Swin Transformer captures long-range dependencies and efficiently processes complex images, forming the backbone of the feature extraction process. The EFRB improved spatial granularity through depthwise convolutions, that focus on spatial features independently across each channel, and pointwise convolutions to improve channel representation. The proposed model used residual connections to enable deeper networks to overcome vanishing gradient problem. Results and discussion: The training process is optimized using population-based feature selection, resulting in robust performance. The network is trained on a dataset split into 80% training and 20% testing, with a learning rate of 1e-3, batch size of 16, and 30 epochs. Evaluation results show that the model achieves an accuracy of 98%, with precision, recall, and F1-scores as 0.97, 0.99, and 0.98 for crack detection, respectively. These results show the effectiveness of the proposed architecture for real-world crack detection applications in structural monitoring.