Visual Saliency Detection Via Tensor Decomposition

Visual attention, also called visual saliency, is the rapid detection of interest regions among massive amount of spatial-temporal visual data. Visual attention, which serves as the reliable and efficient selection of visual information, is the pivot of human vision system. The paper deals with the computational models of bottom-up visual saliency. We focus on the possible explanation of visual saliency and the relation between redundancy and saliency. Therefore, our work could be summarized into two parts:Visual Saliency via Incremental Coding Length: The model re-considers the saliency as the term of coding cost, namely the coding length. To be more precise, based on the spatial-temporal center surround architecture, we model the saliency as the incremental coding length, corresponding to the conditional entropy in the information theory. As solutions towards the model, we proposed two different coding strategies, i.e. Gaussian Conditional Entropy and Sparse Coding. Experiments on salient regions detection and eye tracking data indicate the robustness and efficiency of our method. The work is presented in ACCV 2009 and ICIP 2009.Rank-Sparsity Tensor Decomposition and its application in Visual Saliency: We propose the Rank-Sparsity Tensor Decomposition algorithm. The method could automatically explore the low-dimensional structure of the tensor data, seeking optimal dimension and basis for each mode and separating the irregular patterns. We also develop the saliency detection algorithm via this decomposition. The method is applicable in a wide range of computer vision tasks, including visual attention, foreground segmentation, image denoising and face analysis. The experiments demonstrate promising results.