Hyperparameter tuning of convolutional neural networks for building construction image classification

Deep Learning models have important applications in image processing. However, one of the challenges in this field is the definition of hyperparameters. Thus, the objective of this work is to propose a rigorous methodology for hyperparameter tuning of Convolutional Neural Network for building construction image classification, especially in roofs structure analysis. For this, the HyperTuningSK algorithm was developed, intended to create recommendation rankings for two hyperparameters: learning rate and optimizer. The approach uses concepts from the statistical design of experiments, such as Analysis of Variance and the Scott-Knott clustering algorithm. In addition, the adopted database includes images of inspections on buildings roofs made with unmanned aerial vehicles. The images are divided into two classes: (i) roofs with clean gutters and (ii) roofs with dirty gutters. The methods recommended by the HyperTuningSK algorithm achieved good results in comparison to the hyperparameters adopted in the literature. In this respect, adagrad015 achieved the highest average values of accuracy in the validation (100%) and testing steps (90%) for Convolutional Neural Network architecture with 12 layers. In addition, the hyperparameters recommended by the HyperTuningSK algorithm achieved the best test results for other two literature architectures: Densenet121 (85.7%) and VGG16 (84.4%).