Unsupervised Deep Learning for Online Foreground Segmentation Exploiting Low-Rank and Sparse Priors

This paper proposes a simple approach to unsupervised deep learning for foreground object segmentation. Robust principal component analysis (RPCA) can achieve background subtraction by minimizing nuclear and
$\ell_{1}$
norms to exploit the prior knowledge about spatio-temporal sparseness and low-rankness of the foreground objects and background scene. With a combination of these norms as a loss function, the proposed method trains a U-Net-based model so as to encode and decode the sparse foreground objects for a batch of input images with a low-rank background. Once the model has learned enough features common to the foreground objects, it has the potential to detect them from any single image regardless of the low-rankness and sparseness. The proposed model performs online object segmentation with much less computational expense than that of RPCA. These advantages over RPCA are demonstrated with background subtraction in video surveillance. It is also shown experimentally that the present method can build up a well-generalized cell nuclei segmentation model from only a few dozen of unannotated training images.