Building Infrastructure Classification with Hybrid Convolutional Neural Network Architecture

Since the modernization of structural and architectural building construction, it is now a developing research field for automated building infrastructure categorization and detection for both computer and architectural researchers for the intelligence system. In this research, various types of building infrastructure are classified, such as mosques, temples, brick houses, tin shed houses, and so on. Despite the lack of a standard dataset in this research area, a new dataset called the building infrastructure image dataset is introduced. For the image classification, two types of convolutional neural networks (CNN) architecture are proposed, the first one is sequential CNN and another one is hybrid CNN. The performance of these two architectures is evaluated on the proposed dataset. Again, the proposed architecture’s performance is compared with popular transfer learning models with or without pre-trained weights. With the pre-trained weights, the transfer learning models perform better than the ones without pre-trained weights, which is experimented on in the proposed dataset. The proposed hybrid CNN model outperforms the sequential CNN model as well as popular transfer learning models such as MobileNet and DenseNet without pre-trained weights on the proposed dataset.