An online algorithm for separating sparse and low-dimensional signal sequences from their sum, and its applications in video processing
In signal processing, ``low-rank + sparse'' is an important assumption when separating two signals from their sum. Many applications, e.g., video foreground/background separation are well-formulated by this assumption. In this work, with the ``low-rank + sparse'' assumption, we design and evaluate an online algorithm, called practical recursive projected compressive sensing (prac-ReProCS) for recovering a time sequence of sparse vectors St and a time sequence of dense vectors Lt from their sum, Mt = St + Lt, when the Lt's lie in a slowly changing low-dimensional subspace of the full space.
In the first part of this work (Chapter 1-5), we study and discuss the prac-ReProCS algorithm, the practical version of the original ReProCS algorithm. We apply prac-ReProCS to a key application -- video layering, where the goal is to separate a video sequence into a slowly changing background sequence and a sparse foreground sequence that consists of one or more moving regions/objects on-the-fly. Via experiments we show that prac-ReProCS has significantly better performance compared with other state-of-the-art robust-pca methods when applied to video foreground-background separation.
In the second part of this work (Chapter 6), we study the problem of video denoising. We apply prac-ReProCS to video denoising as a preprocessing step. We develop a novel approach to video denoising that is based on the idea that many noisy or corrupted videos can be split into three parts -- the ``low-rank laye'', the ``sparse layer'' and a small residual which is small and bounded. We show using extensive experiments, layering-then-denoising is effective, especially for long videos with small-sized images that those corrupted by general large variance noise or by large sparse noise, e.g., salt-and-pepper noise.
In the last part of this work (Chapter 7), we discuss an independent problem called logo detection and propose a future research direction where prac-ReProCS can be combined with deep learning solutions.