*3.9. The AdaBoost Classifier*

The AdaBoost classifier is another tool used by machine learning algorithms. It is one of the basic algorithms for boosting, which is a method by which a large number of weak classifiers can be used to get one better. In this case, a weak classifier is one that is relatively simple and can classify test data with an efficiency of more than 50%. The classifier works in such a way that in subsequent iterations it trains and then measures the error of all available weak classifiers. In each subsequent iteration the importance of badly classified observations is increased, so that the classifiers pay more attention to them. The algorithm can significantly improve the quality of classification, but this improvement is only observed when weak classifiers are used as components. When more complex classifiers are used, the use of the AdaBoost algorithm does not lead to a significant increase in effectiveness [93].

Cord and Chambon [94] used the AdaBoost classifier to classify defects on road surfaces. The whole method is based on an appropriate selection of linear and non-linear filters depending on the texture of the analyzed image. In this case, the task of the AdaBoost classifier is to choose the most optimal set of filters to extract cracks in a particular case. A number of textural descriptors are used to learn the classifier. The obtained results were compared with the results of the methods already described in the literature and high efficiency of the developed method was noticed. Prasanna et al. [95] developed an automatic method for the cracks detection on bridge deck surfaces, which they called as the STRUM. The use of machine learning algorithms allowed to eliminate the need for manual definition and tuning of threshold parameters. The developed method uses the robust curve fitting to locate potential cracked areas. Three types of classifiers were used within the algorithm, i.e., the AdaBoost, support vector machines and random forest. The procedure allows to develop the crack density maps in the analyzed area. The accuracy of the algorithm was estimated at about 90%.
