Accurate structural crack detection using NestedUNet from drone and handheld camera images

초록

Crack detection has been significant in structural health monitoring and maintenance of concrete facilities. Many efforts have been made to automate the crack detection process using convolutional neural networks (CNNs). However, many previous CNN architectures and datasets have limitations that require improvements to reduce training time, GPU requirements, and model generalization capabilities. To mitigate these problems, we have developed a new dataset with different types of complex cracks utilizing a handheld camera and drone in different light, environmental conditions, and varying backgrounds. At the same time, we presented an enhanced CNN encoder-decoder architecture that achieves crack segmentation with higher accuracy with fewer parameters. To cope with spatial information loss effectively, we used nested skip connections in the encoder part, along with the EfficientNet-B7 encoder, for speed, accuracy enhancement, and parameter minimization. In the decoder part, residual convolution blocks are employed to alleviate the vanishing gradient problem and aid spatial information recovery. In addition, we added the convolutional block attention module (CBAM) within our architecture to enhance feature extraction by emphasizing crack details over the background. Experimental results demonstrate that our proposed method achieves a 96.15 % accuracy, 85.78 % precision, 82.23 % recall, and 84.00 % F1-score on our custom dataset. Comparisons with other models show that our model outperforms them on all evaluation matrices while having fewer parameters.

키워드