Image quality enhancement and evaluation based on generative adversarial network model for infrastructure inspection using UAV

Abstract

Infrastructure safety inspections based on unmanned aerial vehicles (UAVs) are evaluated to be superior in safety, economic efficiency, and accuracy of inspection results compared to manpower-based inspection methods. In particular, since high-resolution images taken by UAV are directly used to identify and quantify structural defects on the exterior of the facility, the quality of the image has a great impact on the results. However, dynamic blur(motion blur) caused by various external factors such as vibration, low illumination, and wind that occur directly or indirectly in unmanned vehicles can significantly reduce the quality of the image, resulting in inaccurate defect detection or quantified results. Therefore, to solve this problem, this thesis proposes a new image quality evaluation method to evaluate the degree of blur of images acquired from UAV during facility inspection. The proposed methods through the creation of local blur map can effectively classify large image datasets into high-quality and low-quality images. In addition, this thesis propose a deblurring algorithm using Generative Adversarial Network (GAN) to eliminate the motion blur effect of image datasets classified as low-quality. The deblurring method through the GAN model is one of the image restoring methods using the generation model, and operates by modifying the blurred artifacts in the image to generate a clear image. The GAN-based deblur network proposed in this thesis is differentiated from existing studies in that it consists of deblur learning and blur learning modules to eliminate motion blur effects, and generates blur images required for learning of generative model more realistically. Finally, the deblurred UAV image is evaluated through the proposed quality assessment algorithm. In addition, verification experiments based on a publicly available dataset using Full-Reference (FR) evaluation methods demonstrated superior performance compared to existing deblurring algorithms. Furthermore, in surface crack detection experiments utilizing an object detection model (YOLOv7), it was confirmed that utilizing deblurred images yielded better detection results.