Artefacts detection and removal in remote optical imageries using artificial intelligence

• Main Author: You, Zongtao
• Other Authors: Lu Yilong
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/158868
• Institution: Nanyang Technological University

Satellite images have increasingly been used in many different fields, such as detecting and locating ground information, and are used to support fields like urban planning, navigation systems and disaster monitoring. The main problem with satellite images are artefacts or other objects, such as clouds and cloud shadows, that appear in the taken image. These are difficult to detect and remove with increasing image resolution. In this dissertation, both artefact detection and removal are studied. Images are a combination of pixels with different values, and finding a suitable pixel value threshold will improve the ability to detect artefacts. Image inpainting is required after artefact removals to recover the image. In this dissertation two popular image inpainting methods are studied. EdgeConnect uses a two-step process to fix an image: first it generates an edge map of the broken image, and then completes the holes with the help of the generated edges. The Learnable Bidirectional Attention Maps (LBAM) algorithm leverages partial convolution and attention map, and it focuses on completing irregular holes rather than regenerating a whole image. Overall, both methods achieve acceptable results on satellite images that feature irregular holes. LBAM performs slightly better than EdgeConnect because it does not rely on the edge map. It is hard to properly generate the accurate edges of the buildings beneath the artefacts.