The proliferation of camera equipped devices, such as netbooks, smartphones and game stations, has led to a significant increase in the production of visual content. This visual information could be used for understanding the environment and offering a natural interface between the users and their surroundings. However, the massive amounts of data and the high computational cost associated with them, encumbers the transfer of sophisticated vision algorithms to real life systems, especially ones that exhibit resource limitations such as restrictions in available memory, processing power and bandwidth. One approach for tackling these issues is to generate compact and descriptive representations of image data by exploiting inherent redundancies. We propose the investigation of dimensionality reduction and sparse representations in order to accomplish this task. In dimensionality reduction, the aim is to reduce the dimensions of the space where image data reside in order to allow resource constrained systems to handle them and, ideally, provide a more insightful description. This goal is achieved by exploiting the inherent redundancies that many classes of images, such as faces under different illumination conditions and objects from different viewpoints, exhibit. We explore the description of natural images by low dimensional non-linear models called image manifolds and investigate the performance of computer vision tasks such as recognition and classification using these low dimensional models. In addition to dimensionality reduction, we study a novel approach in representing images as a sparse linear combination of dictionary examples. We investigate how sparse image representations can be used for a variety of tasks including low level image modeling and higher level semantic information extraction. Using tools from dimensionality reduction and sparse representation, we propose the application of these methods in three hierarchical image layers, namely low-level features, mid-level structures and high-level attributes. Low level features are image descriptors that can be extracted directly from the raw image pixels and include pixel intensities, histograms, and gradients. In the first part of this work, we explore how various techniques in dimensionality reduction, ranging from traditional image compression to the recently proposed Random Projections method, affect the performance of computer vision algorithms such as face detection and face recognition. In addition, we discuss a method that is able to increase the spatial resolution of a single image, without using any training examples, according to the sparse representations framework. In the second part, we explore mid-level structures, including image manifolds and sparse models, produced by abstracting information from low-level features and offer compact modeling of high dimensional data. We propose novel techniques for generating more descriptive image representations and investigate their application in face recognition and object tracking. In the third part of this work, we propose the investigation of a novel framework for representing the semantic contents of images. This framework employs high level semantic attributes that aim to bridge the gap between the visual information of an image and its textual description by utilizing low level features and mid level structures. This innovative paradigm offers revolutionary possibilities including recognizing the category of an object from purely textual information without providing any explicit visual example.
Library of Congress Subject Headings
Computer vision--Data processing; Image compression; Pattern recognition systems--Mathematics; Machine learning; Ubiquitous computing
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Tsagkatakis, Grigorios, "Dimensionality reduction and sparse representations in computer vision" (2011). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus