采用非线性块稀疏字典选择的视频总结

The current video summarization based on sparse representation does not fully consider the nonlinear relationship between video frames and the block sparsity of key-frames. In this paper, a video summarization scheme with nonlinear block sparse dictionary selection is proposed. The nonlinearity among video frames is considered, and the original video samples are mapped to a very high-dimensional space via a kernel function, which makes the linearly inseparable samples become separable ones to transform the nonlinear cases to linear, thus a nonlinear sparse dictionary selection model is established. The video frames are divided into frame blocks to introduce the block sparsity of key-frames, where the frame contents are similar, and a nonlinear block sparse dictionary selection model is further established to extract key-frame blocks. A kernelized simultaneous block-orthogonal matching pursuit (KSBOMP) method is designed to optimize the proposed model. Experimental analysis for the benchmark video dataset indicates that KSBOMP method can significantly improve the summarization performance in terms of F-score with a low computation complexity, which verifies the effectiveness of simultaneously using nonlinearity and block sparsity.