LRTA-SP: Low-Rank Tensor Approximation With Saliency Prior for Small Target Detection in Infrared Videos

Existing approaches still face issues, such as the lack of spatial–temporal information and target prior information, as well as low detection efficiency when dealing with small infrared (IR) target detection tasks under heterogeneous backgrounds. To address the aforementioned issues, this article presents a low-rank tensor approximation with saliency prior (LRTA-SP) approach, where the holistic spatiotemporal tensor model is constructed by combining spatiotemporal related prior information of IR videos with target saliency prior information. First, benefiting by the subspace optimization theory, the target and the background can be separated into a sparse and a low-rank component, respectively. Crucially, the proposed LRTA-SP optimization approach updates the background tensor on a tangent space to accelerate the above optimization process, greatly reducing the computational complexity of low-rank projection while improving the detection efficiency of our model. Also, applying a multirank constrain in low-rank regularization term helps to adaptively preserve important information in the frequency domain. Furthermore, a prior weight tensor containing target saliency information is provided in sparse regularization term to preserve contextual information during the optimization process. Finally, an alternating projection-based algorithm framework is designed to robustly separate sparse targets and low-rank backgrounds. The effectiveness and superiority, especially the detection efficiency, of the proposed LRTA-SP technology to similar detection technologies are validated on six real IR videos under various scenarios.