Machine Learning Hyperparameter Recommendation Methods in Building Image Classification
Building Construction, Data Augmentation, Hyperparameter Tuning, Image Classification, Machine Learning.
This work proposes rigorous methods for hyperparameter tuning of machine learning for classifying images in construction. For this, computational methods called HyperTuningSK and AutoHyperTuningSK are proposed. These algorithms use statistical techniques for recommending hyperparameters, such as Analysis of Variance and the Scott-Knott clustering algorithm. The approach uses statistical experimental design concepts, such as analysis of variance and the Scott-Knott clustering algorithm. In addition, four case studies were used: façade vegetation detection, gutter integrity detection, machinery classification, and crack classification. The results showed that the hyperparameters affect the performance of image classification. It is also worth noting that the hyperparameter configurations adjusted by HyperTuningSK resulted in different recommendations depending on the neural architecture used. In this sense, the adagrad025 combination achieved a HyperScore = AAA for the Densenet121 architecture. The results of the data augmentation analysis show that two transformations were the most recommended by HyperTuningSK: width shift and height shift. Moreover, the AutoHyperTuningSK algorithm recommended an adagrad optimizer and a learning rate of 0.02220 in the experiments for the fourth case study. This combination achieved a maximum accuracy of 99.48%, that is, the correct classification of 3,979 images (4,000 in total) in the test dataset. The results for tuning data augmentation hyperparameters also confirm the efficiency of the proposed approach using the AutoHyperTuningSK-DA algorithm. In this regard, the recommended combination achieved an average accuracy of 99.2% in the test experiments.