Increasing the robustness of material-specific deep learning models for crack detection across different materials

摘要

Infrastructure defect detection solutions based on computer vision have recently emerged as powerful tools with applications in both traditional inspection practices, as well as robotic inspections. These applications involve the collection of images from a wide range of infrastructure systems with heterogeneous characteristics such as conditions, materials, surface appearances and textures. Consequently, defect detection models need to be sufficiently robust to accommodate this type of heterogeneity. Existing image-based crack detection literature almost entirely focuses on models tailored to crack detection in either concrete or asphalt surfaces with prior knowledge of the material involved and studies on crack detection in more than one material are needed for truly automated inspection systems. This paper focuses on the adaptability of deep learning-based crack detection models across common construction materials. To investigate this problem, a residual convolutional neural network architecture was trained and tested on two separate concrete and asphalt crack image data sets and compared with existing baselines. These tests demonstrated that the change of material significantly reduces crack detection accuracy of a tailored model. In response, three domain adaptation techniques, namely joint training, sequential training, and ensemble learning are proposed and implemented to develop robust crack detection models that work on both datasets regardless of the material environment. Results demonstrate that the proposed techniques are able to successfully produce accuracies comparable to those of the material-specific models, without prior knowledge of the material.