With widespread availability of low-cost devices capable of photo shooting and high-volume video recording, we are facing explosion of both image and video data. The sheer volume of such visual data poses both challenges and opportunities in machine learning and computer vision research. In image classification, most of previous research has focused on small to mediumscale data sets, containing objects from dozens of categories. However, we could easily access images spreading thousands of categories. Unfortunately, despite the well-known advantages and recent advancements of multi-class classification techniques in machine learning, complexity concerns have driven most research on such super large-scale data set back to simple methods such as nearest neighbor search, one-vs-one or one-vs-rest approach. However, facing image classification problem with such huge task space, it is no surprise that these classical algorithms, often favored for their simplicity, will be brought to their knees not only because of the training time and storage cost they incur, but also because of the conceptual awkwardness of such algorithms in massive multi-class paradigms. Therefore, it is our goal to directly address the bigness of image data, not only the large number of training images and high-dimensional image features, but also the large task space. Specifically, we present algorithms capable of efficiently and effectively training classifiers that could differentiate tens of thousands of image classes. Similar to images, one of the major difficulties in video analysis is also the huge amount of data, in the sense that videos could be hours long or even endless. However, it is often true that only a small portion of video contains important information. Consequently, algorithms that could automatically detect unusual events within streaming or archival video would significantly improve the efficiency of video analysis and save valuable human attention for only the most salient contents. Moreover, given lengthy recorded videos, such as those captured by digital cameras on mobile phones, or surveillance cameras, most users do not have the time or energy to edit the video such that only the most salient and interesting part of the original video is kept. To this end, we also develop algorithm for automatic video summarization, without human intervention. Finally, we further extend our research on video summarization into a supervised formulation, where users are asked to generate summaries for a subset of a class of videos of similar nature. Given such manually generated summaries, our algorithm learns the preferred storyline within the given class of videos, and automatically generates summaries for the rest of videos in the class, capturing the similar storyline as in those manually summarized videos.

Reducing interference noise in a noisy speech recording has been a challenging task for many years yet has a variety of applications, for example, in handsfree mobile communications, in speech recognition, and in hearing aids. Traditional single-channel noise reduction schemes, such as Wiener filtering, do not work satisfactorily in the presence of non-stationary background noise. Alternatively, supervised approaches, where the noise type is known in advance, lead to higher-quality enhanced speech signals. This dissertation proposes supervised and unsupervised single-channel noise reduction algorithms. We consider two classes of methods for this purpose: approaches based on nonnegative matrix factorization (NMF) and methods based on hidden Markov models (HMM).  The contributions of this dissertation can be divided into three main (overlapping) parts. First, we propose NMF-based enhancement approaches that use temporal dependencies of the speech signals. In a standard NMF, the important temporal correlations between consecutive short-time frames are ignored. We propose both continuous and discrete state-space nonnegative dynamical models. These approaches are used to describe the dynamics of the NMF coefficients or activations. We derive optimal minimum mean squared error (MMSE) or linear MMSE estimates of the speech signal using the probabilistic formulations of NMF. Our experiments show that using temporal dynamics in the NMF-based denoising systems improves the performance greatly. Additionally, this dissertation proposes an approach to learn the noise basis matrix online from the noisy observations. This relaxes the assumption of an a-priori specified noise type and enables us to use the NMF-based denoising method in an unsupervised manner. Our experiments show that the proposed approach with online noise basis learning considerably outperforms state-of-the-art methods in different noise conditions.  Second, this thesis proposes two methods for NMF-based separation of sources with similar dictionaries. We suggest a nonnegative HMM (NHMM) for babble noise that is derived from a speech HMM. In this approach, speech and babble signals share the same basis vectors, whereas the activation of the basis vectors are different for the two signals over time. We derive an MMSE estimator for the clean speech signal using the proposed NHMM. The objective evaluations and performed subjective listening test show that the proposed babble model and the final noise reduction algorithm outperform the conventional methods noticeably. Moreover, the dissertation proposes another solution to separate a desired source from a mixture with arbitrarily low artifacts.  Third, an HMM-based algorithm to enhance the speech spectra using super-Gaussian priors is proposed. Our experiments show that speech discrete Fourier transform (DFT) coefficients have super-Gaussian rather than Gaussian distributions even if we limit the speech data to come from a specific phoneme. We derive a new MMSE estimator for the speech spectra that uses super-Gaussian priors. The results of our evaluations using the developed noise reduction algorithm support the super-Gaussianity hypothesis. / <p>QC 20130916</p>