Tuning of data augmentation hyperparameters in deep learning to building construction image classification with small datasets

Abstract

Deep Learning methods have important applications in the building construction image classification field. One challenge of this application is Convolutional Neural Networks adoption in a small datasets. This paper proposes a rigorous methodology for tuning of Data Augmentation hyperparameters in Deep Learning to building construction image classification, especially to vegetation recognition in facades and roofs structure analysis. In order to do that, Logistic Regression models were used to analyze the performance of Convolutional Neural Networks trained from 128 combinations of transformations in the images. Experiments were carried out with three architectures of Deep Learning from the literature using the Keras library. The results show that the recommended configuration (Height Shift Range = 0.2; Width Shift Range = 0.2; Zoom Range =0.2) reached an accuracy of in the test step of first case study. In addition, the hyperparameters recommended by proposed method also achieved the best test results for second case study: .

Aplicação de Aprendizado Profundo na Classificação de Imagens de Patologias da Construção Civil: Análise de Data Augmentation para Bancos de Dados Desbalanceados

“…O ramo das Redes Neurais Artificiais (RNAs) se destaca entre as aplicac ¸ões de aprendizado de máquina em virtude da alta precisão obtida no reconhecimento de padrões [2], [43], [55]. Diante disso, os seus resultados são a consequência da arquitetura de aprendizagem e processamento representada pela Figura 1 baseado na similaridade com a estrutura de neurônios biológicos [64], [15].…”

“…No entanto, apesar da ampla aplicac ¸ão em segmentos como danos estruturais [67], otimizac ¸ão na utilizac ¸ão de materiais [36], custo total de produc ¸ão [38], índice de produtividade da construc ¸ão [16] e seguranc ¸a na construc ¸ão de barragens [31], este tipo de treinamento das CNNs em sua melhor perfomance possui desafios para a aplicac ¸ão na construc ¸ão civil [43], [38], [44], [57]. Isso porque, a coleta de imagens no formato tradicional é efetuada de forma manual, dificultando assim a alta escalabilidade de fotografias necessárias [46].…”

Tuning of Data Augmentation Hyperparameters to Covid-19 Detection in X-Ray Images with Deep Learning

“…Following this aspect of defining good conditions for Deep Learning experiments, an important area is Data Augmentation [18][19][20][21]. This technique is used to reduce overfitting, generating artificial training images through random transformations [22].…”

“…This technique is used to reduce overfitting, generating artificial training images through random transformations [22]. The application of this method becomes important especially in cases where the number of samples for training is relatively small [18,21,22]. In this context, most of the studies on the application of Deep Learning in the diagnosis of Covid-19 via images are found, since they perform experiments with databases with few examples [16,23,24].…”

“…Furthermore, it is suggested to analyze with other databases, in order to verify the model performance beyond the images initially used and to validate the accuracy hypothesis. Finally, it is recommended to use other statistical measures for analysis (standard deviation, median and variance) and apply other methodologies for hyperparameter tuning [21,39].…”