TY - JOUR
T1 - A computationally efficient crack detection approach based on deep learning assisted by stockwell transform and linear discriminant analysis
AU - Nguyen, A.
AU - Nguyen, C. L.
AU - Gharehbaghi, V.
AU - Perera, R.
AU - Brown, J.
AU - Yu, Yang
AU - Kalbkhani, H.
PY - 2022
Y1 - 2022
N2 - This paper presents SpeedyNet, a computationally efficient crack detection method. Rather than using a computationally demanding convolutional neural network (CNN), this approach made use of a simple neural network with a shallow architecture augmented by a 2D Stockwell transform for feature transformation and linear discriminant analysis for feature reduction. The approach was employed to classify images with minute cracks under three simulated noisy conditions. Using time–frequency image transformation, feature conditioning and a fast deep learning-based classifier, this method performed better in terms of speed, accuracy and robustness compared to other image classifiers. The performance of SpeedyNet was compared to that of two popular pre-trained CNN models, Xception and GoogleNet, and the results demonstrated that SpeedyNet was superior in both classification accuracy and computational speed. A synthetic efficiency index was then defined for further assessment. Compared to GoogleNet and the Xception models, SpeedyNet enhanced classification efficiency at least sevenfold. Furthermore, SpeedyNet's reliability was demonstrated by its robustness and stability when faced with network parameter and input image uncertainties including batch size, repeatability, data size and image dimensions.
AB - This paper presents SpeedyNet, a computationally efficient crack detection method. Rather than using a computationally demanding convolutional neural network (CNN), this approach made use of a simple neural network with a shallow architecture augmented by a 2D Stockwell transform for feature transformation and linear discriminant analysis for feature reduction. The approach was employed to classify images with minute cracks under three simulated noisy conditions. Using time–frequency image transformation, feature conditioning and a fast deep learning-based classifier, this method performed better in terms of speed, accuracy and robustness compared to other image classifiers. The performance of SpeedyNet was compared to that of two popular pre-trained CNN models, Xception and GoogleNet, and the results demonstrated that SpeedyNet was superior in both classification accuracy and computational speed. A synthetic efficiency index was then defined for further assessment. Compared to GoogleNet and the Xception models, SpeedyNet enhanced classification efficiency at least sevenfold. Furthermore, SpeedyNet's reliability was demonstrated by its robustness and stability when faced with network parameter and input image uncertainties including batch size, repeatability, data size and image dimensions.
UR - https://hdl.handle.net/1959.7/uws:78213
U2 - 10.1016/j.istruc.2022.09.107
DO - 10.1016/j.istruc.2022.09.107
M3 - Article
SN - 2352-0124
VL - 45
SP - 1962
EP - 1970
JO - Structures
JF - Structures
ER -