BSVOS: Background Interference Suppression Strategy for Satellite Video Multiobject Segmentation

Abstract

Video satellites provide dynamic real-time monitoring of hotspot areas and objects by continuously imaging objects within a specified sequence, providing dynamic information over large areas. Satellite video object segmentation aims to separate foreground and background continuously. Compared with satellite video single-object segmentation, satellite video multi-object segmentation can separate multiple foreground objects of interest in the scene simultaneously. So far, the research on multi-object segmentation mainly focuses on extracting deeper informative features. However, this does not cope well with the problems of weak object texture information, foreground-background disproportion, and small object scale in satellite videos. Therefore, this paper proposes an end-to-end semi-supervised segmentation network named background interference suppression strategy for satellite video multi-object segmentation (BSVOS), which is the first research about satellite video multi-object segmentation. First, spatial and temporal matching templates integrate spatio-temporal information to enhance texture features within a sequence. Second, a novel matching strategy is also proposed to suppress complex background interference in satellite video. Finally, both intuitive and quantitative experimental results indicate that our proposed method outperforms other multi-object segmentation models.