A Novel Paradigm for Thick Cloud Removal in Aerial Videos

The presence of thick clouds has a significant impact on aerial image processing and application. It is difficult to obtain satisfactory results by using only a single image to compensate the cloud region content. How to use video information to remove the interference of thick clouds and maintain the original appearance of the scene is a seemingly pathological problem. Therefore, we propose a novel thick cloud removal paradigm combining semi-supervised segmentation and optical flow estimation model. Firstly, the first frame of the video is labeled at pixel level, and the whole video is automatically segmented in the cloud region by using the semi-supervised video object segmentation model, and then the optical flow of ground movement is estimated by the optical flow estimation model. In order to eliminate the pixel differences of the same scenes in different frames, a pixel fusion strategy is designed. We use optical flow to fill the cloud area with real information and get preliminary results. Compared with remote sensing image and single image cloud removal methods, the processing results of proposed method are more natural and reasonable.