Image Quality Assessment for UAV-based Infrastructure Inspection: An AI Pre-selection Strategy Ya-Li Lin(a*), Guan-Chin Su(a), Lai-Han Zou(a), Chao-Hung Lin(b), Jiann-Yeou Rau(b), Wei-Shen Lai(c), Chih-Chao Hu(c)
(a) Student, Department of Geomatics, National Cheng-Kung University, Taiwan
* alecfree2[at]gmail.com
(b) Professor, Department of Geomatics, National Cheng-Kung University, Taiwan
(c) Researcher, Transportation Engineering Division, Institute of Transportation, Taiwan
Abstract
Image quality assessment is a fundamental step for image-based technologies such as deep learning detection, 3D reconstruction, and deformation monitoring. In UAV-based infrastructure inspection, image quality often declines because of the platform vibration, unstable lighting, and motion blur, which affects the accuracy and consistency of downstream analysis. Current image selection rely on manual and visual inspection, which is generally labor-intensive and qualtiy inconsistent. Although no-reference image quality assessment (IQA) has gained increasing attention in recent years, existing methods rely on subjectively labeled datasets which lack the objectivity and interpretability required in real-world applications. This study proposes a reference-free IQA framework designed for UAV imagery, and introduces a structural similarity index enhanced with a CLIP encoder (CSSIM) to address pixel misalignment caused by image cropping. A standardized and objective data construction pipeline is developed to support the generation of image quality maps (IQMs) using a Swin-Unet model. The Swin-Unet network architecture enables fast, full-image inference, providing consistent and efficient assessment of high-resolution UAV images and making IQA feasible for field deployment. Furthermore, the generated IQMs are used in an AI image selection strategy for UAV-based infrastructure inspection, allowing automatic removal of low-quality images prior to tasks such as bridge geometry reconstruction and structural recognition. The proposed method has been successfully applied in real UAV bridge inspection scenarios, significantly improving model prediction accuracy and showcasing its value in scalable, quality-aware infrastructure monitoring.
Keywords: UAV, image quality accessment, deep learning, infrastructure monitoring
