*3.6. Artificial Neural Networks*

Neural networks are a very complex family of mathematical structures whose software or hardware models perform calculations or process signals through rows of processing elements called as the artificial neurons. The artificial neurons perform a basic operation on their input. There are different types of artificial neural networks, but their common feature is that their structure consists of neurons connected by synapses. Weights, or numerical values, are associated with synapses, whose interpretation depends on the model [74].

Lee and Lee [75] presented an integrated system consisting of three types of neural networks (image-based, histogram-based and proximity-based) to classify cracks on the concrete pavement surface. The system was validated on 124 actual pavement images. As a result of the analysis, it was concluded that the proximity-based neural network, despite the smallest computational requirements, is characterized by the highest accuracy in searching for the cracking patterns, which was estimated at about 95%. Bray et al. [76] developed a two-stage process of classifying cracks on the surface of concrete road surfaces using the artificial neural networks. In the first stage, based on the density and histogram, the artificial neural network divided into images with cracks and images without cracks. In the second stage, another neural network, after image segmentation, determined the type of cracks. The obtained results indicated 100% accuracy of the system in classifying images with cracks and 82% accuracy in classifying images without cracks. While the developed system gave good results in terms of cracks detection, it did not perform well in classifying the type of crack. Moon and Kim [77] developed an intelligent, automatic system for the cracks detection on concrete surfaces. The algorithm consisted of two main stages. In the first step, cracks were extracted from the digital image using filtering operations, the improved subtraction method and morphological operations. In the second stage, the cracks detection was performed by means of a backpropagation neural network. The whole algorithm was trained on 105 images of cracked concrete surfaces and then validated on another 120 images. The developed algorithm was characterized by a very high recognition rate, amounting to 90% for crack images and 92% for non-crack images, respectively. The authors emphasize the simplicity of the application of the algorithm, which makes it possible to assess the occurrence of cracks by non-expert inspectors. The versatility of programming artificial neural networks made it a very popular technique for detection of the cracking patterns, where further examples of application can be found in [78–81]. An example of the effect of identifying cracks on concrete surfaces with the implementation of artificial neural networks is shown in Figure 8.

**Figure 8.** Results of crack detection by means of the artificial neural network, based on [79].
