Improving crack segmentation with noise-augmented training: model comparison and robustness evaluation on custom dataset

Purpose With the rapid advancement of computer vision and deep learning, crack detection has transitioned from manual inspection to automated approaches. However, challenges such as varying illumination and environmental noise continue to hinder detection accuracy. This study aims to enhance crack segmentation performance and robustness under complex imaging conditions through noise-augmented training and rigorous model comparison. Design/methodology/approach Noise-augmented versions of public benchmark datasets were employed to train selected segmentation models, thereby enhancing their robustness to illumination variations and noise interference. To evaluate model generalization, a challenging dataset containing 434 images featuring diverse infrastructure types and camera angles was constructed. Two deep learning frameworks, DeepLabv3+ and SegNet, were implemented with various pre-trained backbones, resulting in seven distinct architectures, such as DeepLabv3+Inception-ResNet-v2, for comparative performance analysis. Findings Models trained on noise-augmented datasets exhibited notable improvements in Mean Intersection over Union (MIoU) and F1-score compared with their non-augmented counterparts. Specifically, the DeepLabv3+Inception-ResNet-v2 model achieved the most significant progress and the best overall performance, demonstrating respective increases of 0.7% in Accuracy, 3.2% in Recall, 15.3% in Precision, 4.4% in F1-score and 5.0% in MIoU on the test set. Furthermore, evaluation on the 434-image dataset confirmed the model's high robustness. Originality/value These findings indicate that the proposed network, DeepLabv3+Inception-ResNet-v2, has strong potential for crack segmentation tasks in basic infrastructure, suggesting its applicability in real-world engineering scenarios.