Saliency Detection With Features From Compressed HEVC

In this paper, a saliency detection algorithm with features from compressed high-efficiency video coding (HEVC) is proposed. The proposed algorithm consists of three parts: static saliency detection, dynamic saliency detection, and competitive fusion. Static features are generated by downsampling and discrete cosine transform, and dynamic features are extracted from compressed HEVC, specifically motion vector. A Gaussian kernel is used to extract the data structure in static feature maps. For dynamic feature map, a coding unit depth and bits combined mask is designed to filter out the dynamic background. Finally, competitive fusion is designed to adaptively fuse the static and dynamic saliency maps. Experimental results show that the proposed method is superior to classic methods by up to 0.1223 area under curve gaining and 0.8362 Kullback-Leibler divergence decreasing on average. The average detection speed is 2.3 s per frame.