A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Deep learning and infrared thermography for asphalt pavement crack severity classification
Abstract Deep learning, especially convolutional neural network (CNN), is becoming a popular and powerful tool for crack detection. This work aims to apply deep learning and infrared thermography for asphalt pavement crack severity classification. A dataset of asphalt pavement crack was built in this work, including four levels of crack severity, no crack, low-severity crack (i.e., low crack), medium-severity crack (i.e., medium crack), and high-severity crack (i.e., high crack). This dataset had three types of images, the visible image, infrared image, and the fusion of visible and infrared images (i.e., fusion image). Thirteen typical CNN models were trained and evaluated on the aforementioned dataset for deep learning from scratch, while eight pre-trained CNN models trained by ImageNet were also trained and evaluated for transfer learning. This work investigated the effects of image types on the accuracy of deep learning from scratch and transfer learning, as well as the effects of image types on classifying the levels of crack severity. The results show that the CNN models had the highest accuracy on the fusion image for deep learning from scratch, but the highest accuracy on the visible image for transfer learning. The CNN models performed well on both the no crack and low crack but had different performances on the medium crack and high crack for all three types of images, while misclassification occurred mainly on the medium crack and high crack for all three types of images. EfficientNet-B3 had the highest accuracy on all three types of images for both deep learning from scratch and transfer learning.
Highlights Build a dataset with four crack severity levels and three image types. Study the effect of image types on deep learning from scratch and transfer learning. Study the effect of image types on classifying crack severity levels. Compare the performance of ResNet, DenseNet, and EfficientNet in crack severity classification.
Deep learning and infrared thermography for asphalt pavement crack severity classification
Abstract Deep learning, especially convolutional neural network (CNN), is becoming a popular and powerful tool for crack detection. This work aims to apply deep learning and infrared thermography for asphalt pavement crack severity classification. A dataset of asphalt pavement crack was built in this work, including four levels of crack severity, no crack, low-severity crack (i.e., low crack), medium-severity crack (i.e., medium crack), and high-severity crack (i.e., high crack). This dataset had three types of images, the visible image, infrared image, and the fusion of visible and infrared images (i.e., fusion image). Thirteen typical CNN models were trained and evaluated on the aforementioned dataset for deep learning from scratch, while eight pre-trained CNN models trained by ImageNet were also trained and evaluated for transfer learning. This work investigated the effects of image types on the accuracy of deep learning from scratch and transfer learning, as well as the effects of image types on classifying the levels of crack severity. The results show that the CNN models had the highest accuracy on the fusion image for deep learning from scratch, but the highest accuracy on the visible image for transfer learning. The CNN models performed well on both the no crack and low crack but had different performances on the medium crack and high crack for all three types of images, while misclassification occurred mainly on the medium crack and high crack for all three types of images. EfficientNet-B3 had the highest accuracy on all three types of images for both deep learning from scratch and transfer learning.
Highlights Build a dataset with four crack severity levels and three image types. Study the effect of image types on deep learning from scratch and transfer learning. Study the effect of image types on classifying crack severity levels. Compare the performance of ResNet, DenseNet, and EfficientNet in crack severity classification.
Deep learning and infrared thermography for asphalt pavement crack severity classification
Liu, Fangyu (author) / Liu, Jian (author) / Wang, Linbing (author)
2022-05-22
Article (Journal)
Electronic Resource
English
Asphalt pavement crack detection based on infrared thermography and deep learning
| Taylor & Francis Verlag | 2024
Infrared Thermography for Asphalt Crack Imaging and Automated Detection
| British Library Online Contents | 2004