基于时空特征融合的全景视频显著性检测网络

utomaticly capture video salient region is very important to guide users to watch panoramic video. This paper proposes a saliency detection network for panoramic video based on spatio-temporal features fusion. The network consists of preprocessing module, feature coding module, salient prediction module and output integration module. The preprocessing module solves the boundary stretching and image distortion problem of panoramic video in the equiangular rectangular plane by cube-padding mapping. The feature coding module extracts multi-scale context spatio-temporal features, and realizes the feature fusion process of spatiot-emporal features adaptive weight based on attention mechanism. The saliency prediction module uses Bi-ConvLSTM to model the temporal information between frames, and the salient region of the current frame.is predicted by combining the results of the previous frame and the next frame. The model achieves better accuracy than the existing models on two public datasets, which verifies the effectiveness of the model.