Performance Analysis of Using Feature Fusion for Crack Detection in Images of Historical Buildings

In this paper, three types of feature sets are used for evaluating the performance of a proposed approach for crack detection in images of historical buildings. The feature sets are hand-crafted features, Convolutional Neural Network (CNN) learned features, and fusion of hand-crafted and CNN-learned features. The proposed approach is validated by implementing several Machine Learning (ML) classifiers with applying 3-fold cross validation. Two datasets of crack images are used for developing the feature sets. Experimental results show that both Support Vector Machine (SVM) and stacked ensemble classifiers achieve highest accuracy of 98% for crack detection using the CNN-learned features with dimensionality reduction. The significance of this study is to highlight the impact of different types of feature sets on the performance of the classification process for crack detection.