SGN-90206 International Doctoral Seminar: Signal Processing Journal Club


Sparse representations are one of the most important models in signal/image processing. These models were originally introduced for descriptive tasks such as denoising/regression, but more recently discriminative tasks, classification in particular, have also been addressed in the sparse representation literature. During this talk, I will present our work on exploiting sparse representation for change/anomaly detection, namely, the identification of data departing from the normal (stationary) conditions. We assume that normal data admits a sparse representation w.r.t. a suitable dictionary, thus that they are well approximated in a union of low-dimensional subspaces. Performing change/anomaly detection then corresponds to identifying data that departs from these subspaces, both because of a permanent shift in the data-generating process (change-detection) or because of a sporadic event (anomaly detection). I will illustrate the effectiveness of this approach in two application scenarios. The first consists of an environmental monitoring application, where nodes of a sensor network deployed over a rock face acquires acoustic emissions (bursts). The stream of bursts is then analyzed to detect structural changes which might be associated with macroscopic variation in the monitored phenomenon, eventually leading to a rock collapse. The second consists of an industrial monitoring application, where a scanning electron microscope (SEM) is used to supervise the production of nanofibres. Here, we address the problem of detecting anomalous patterns in the acquired images, which might indicate a degradation in the quality of the produced materials.