Comparative Study of Various Deep Learning Models for Structural Anomaly Detection

Efficient detection and segmentation of wall cracks play a crucial role in building maintenance and construction. However, the implementation of vision transformers (ViT) for crack classification often presents challenges due to its high computational complexity, making it unsuitable for deployment on low-efficiency devices. To address this issue, we propose a novel approach that leverages knowledge distillation (KD) to create a computationally efficient ensemble model comprising a convolutional neural network (CNN) and ViT. In our framework, the teacher model is the ViT, which possesses exceptional classification capabilities, while the student model is a CNN designed to reduce complexity and enhance inference efficiency. By employing KD, we transfer knowledge from the ViT to the CNN, enabling the student model to approximate the performance of the more complex teacher model. This approach reduces the training time and computational requirements without significantly sacrificing classification accuracy. Following the classification stage, we employ a UNet segmentation model on the crack-detected images to accurately identify and delineate the damaged areas within a cracked surface. By analyzing the segmented images, we can calculate essential metrics such as the percentage of crack area and the length of the cracks. These metrics provide valuable insights into the severity of the cracks, facilitating the development of effective strategies for repair and prevention. Experimental results on a diverse dataset demonstrate that our ensemble model achieves competitive crack detection and segmentation performance while maintaining efficiency. The proposed approach not only reduces the complexity associated with ViT deployment but also provides an accurate and comprehensive analysis of wall crack severity. These findings have significant implications for building maintenance and construction industries, enabling proactive measures to mitigate structural damage and ensure safer and more sustainable infrastructure.