Ship Removal from High-Resolution Satellite Images: Dataset and Challenges

The removal of sensitive objects is critical for the secure sharing of remote sensing resources. Manual object removal is labor-intensive, and automated removal falls short, posing challenges in achieving the necessary quality for sensitive object removal. In this paper, we propose an integrated approach, combining state-of-the-art object detection models with advanced image inpainting techniques in a two-stage processing model tailored for detecting and removing sensitive targets in remote sensing images. We validate the pipeline’s performance using ship targets in remote sensing mapping. Specifically, we first curate a ship removal dataset, dubbed RSInpainting-Ships. We then introduce a refined random-mask strategy, guiding the network towards filling the background within the targeted region for removal. Initial experiments reveal three primary factors influencing performance: size, location, and wake. Size relates to the quantity of pixels to be filled and the complexity of the contextual information, location influences gradient changes in background structure, and wake brings in additional texture and determines the performance of ship removal.