Deep Learning–Based Enhancement of Motion Blurred UAV Concrete Crack Images

Abstract

Building façade inspection and maintenance needs to be carried out periodically, and the detection of cracks is a core component of the inspection process. The current inspection procedure is labor-intensive and time-consuming and poses significant safety issues like falling from height. Unmanned aerial vehicles (UAVs) with computer vision techniques represent a promising approach for visual crack inspection on high-rise building façades. One research challenge to achieving automated visual crack inspection is image degradation in the form of motion blur caused by UAVs during image acquisition. Motion blur arises due to excessive vibrations of the UAV platform, and this may adversely affect crack detection. In this paper, a deep learning–based deblurring model based on a generative adversarial network (GAN) is proposed to address this challenge. Further, by recognizing a strong correlation between blurred and sharpened crack images, the idea of using a localized skip connection is introduced. Experimental validation of the proposed deblurring model is carried out by investigating the impact of skip connections on deblurring. The proposed model is also compared against the state-of-the-art deblurring model, and results indicate that the proposed model is able to achieve significant improvements in deblurring performance in terms of both global structure and feature details in crack images.