Page 89 - Read Online

Wang et al. Intell Robot 2022;2(4):391-406 https://dx.doi.org/10.20517/ir.2022.25 Page 395

[33]
Sparse decomposition is different from the traditional feature extraction method. Sparse decomposition
decomposes noisy signals on redundant dictionaries to achieve feature extraction. In the research of bridge
crack identification, Li et al. adopted a self-learning algorithm to extract scale invariant features from 27,471
unlabeled bridge pavement crack images, adopted an improved sparse coding representation to obtain a
feature dictionary, adopted a spatial pyramid pooling method for feature extraction, and then employed a
[34]
linear support vector machine (SVM) classifier for crack identification . Wang et al. collected hundreds of
concrete crack images and proposed a fast detection method for concrete cracks based on L2 sparse
[35]
representation . To suppress the noise disturbances, discrete cosine transformation is applied to extract the
frequency-domain characteristics of the crack and non-crack image regions. The established complete
dictionary was used to quickly calculate its sparse coefficient and effectively select candidate cracks via a
pooling operation. Sparse decomposition can be succinctly expressed as a linear combination of several
bases, more comprehensively and carefully characterize some features covered by a signal and more
effectively separate the signal and noise. However, sparse decomposition involves a vast amount of
computation and computational complexity.

The WT is mainly based on the different distributions in the frequency domain of the noise and signals and
decomposes a noisy signal into multiple scales. Then, the wavelet coefficients belonging to the noise are
removed at each scale, while the wavelet coefficients belonging to the signal are retained and enhanced.
[36]
After wavelet denoising, the signal is reconstructed to achieve effective feature extraction . Nguyen et al.
measured displacement signals with four separate measurement channels through a sensor system, where
[37]
the displacement signals were simultaneously transmitted at a sampling speed of 100 samples per second .
Then, the WT method was applied to the original deflection signals to decompose the signals into elements
and to eliminate interference signals, which extracts the features of multiple cracks under the action of
moving loads and improves the sensitivity and accuracy in the identification process. In the study of the
finite element model of a simply supported beam with a transverse crack, Nigam et al. employed the WT,
which uses the deflected edge as input for identifying the crack location in the beam, and obtained a good
[38]
detection effect . Because WT retains most of the wavelet coefficients of the signal, the image details can be
well preserved after noise reduction. As the abrupt part of the signal will not be damaged during noise
reduction, it has a good denoising effect. However, since different signals are applicable to different wavelet
bases, it is difficult to identify the optimal wavelet base.

The ANFIS [39,40] organically combines fuzzy logic inference and neural networks and performs self-adaptive
learning on fuzzy experience and knowledge while applying reasoning similar to the human brain to
eliminate noise and interference and to extract feature information [41,42] . Bilir et al. applied the results of free
shrinkage tests conducted to determine the length changes and ring tests performed to determine the
restrained drying shrinkage cracks for predicting the crack widths of granulated blast furnace slag fine
aggregate mortars with ANFIS on 456 data and used the replacement ratios, drying time and free shrinkage
length changes as inputs and crack width as output to predict the shrinkage cracking of the mortar types .
[43]
The ANFIS uses the neural network self-learning ability and fuzzy logic reasoning ability to extract fuzzy
rules from the dataset and calculates the optimal parameters of the membership function to adaptively mine
the sensitive features in the data.

With the rapid development and wide application of artificial intelligence and deep learning, an increasing
number of deep learning methods have been applied to concrete bridge crack detection, especially in crack
feature extraction [44,45] . In research on the automatic classification of concrete structure crack damage based
on cascade generalized neural networks, Guo et al. employed 10,000 cracked and uncracked images from
wall images and pavement images with a splitting ratio for datasets of 8:2 and proposed a cascade broad