Abstract:The field of computer vision based civil infrastructure defect detection is constantly evolving with steady advances being made in sensing technologies, hardware, and image processing techniques. Although a number of image processing techniques have been proposed over the years with varying degrees of success, real-world situations (e.g., lack of background illumination, shadow changes) pose a significant challenge to the wide adoption of such techniques, especially for routine analysis. With the emerging application of Unmanned Aerial Vehicles (UAVs) or drones for civil infrastructure inspection and condition monitoring, an added level of complexity is introduced in vision-based crack detection as the process of acquiring images using UAVs is not yet standardized resulting in images of widely varying sizes, resolutions, blurriness, etc. In recent years, Deep Learning, a generalized form of Deep Neural Network (DNN) algorithms that can learn very complex mappings between inputs and outputs directly from the data, has achieved huge success in diverse fields such as automatic speech recognition, image recognition, Natural Language Processing (NLP), drug and materials discovery, etc. However, the large number of hidden neurons and layers used in DNNs result in computationally-intensive matrix and vector computations involving millions of parameters, requiring the use of high-performance computing systems. Also, it is practically impossible to get labeled "big data" samples in many domains to be able to train an entire DNN from scratch. In such situations, the use of a pre-trained deep learning model and fine-tuning it to the novel task at hand with smaller datasets, has shown to be successful across domains. In this paper, we propose the use of pre-trained deep learning models with transfer learning for crack damage detection in UAV images of civil infrastructure. The robustness of the proposed approach is tested on a small set of real-world, complex UAV-sourced infrastructure images not used during training and validation. The results show that the proposed method can rapidly and easily achieve up to 90% accuracy in finding cracks in realistic situations without any augmentation and preprocessing.