Sparse signal processing for machine learning and computer vision
Date
2014
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Signal sparse representation solves inverse problems to find succinct expressions of data samples as a linear combination of a few atoms in the dictionary or codebook. This model has proven effective in image restoration, denoising, inpainting, compression, pattern classification and automatic unsupervised feature learning. Many classical sparse coding algorithms have exorbitant computational complexity in solving the sparse solution, which hinders their applicability in real-world large-scale machine learning and computer vision problems. In this dissertation, we will first present a family of locality-constrained dictionary learning algorithms, which can be seen as a special case of sparse coding. Compared to classical sparse coding, locality-constrained coding has closed-form solution and is much more computationally efficient. In addition, the locality-preserving property enables the newly proposed algorithms to better exploit the geometric structures of data manifold. Experimental results demonstrate that our algorithms are capable of achieving superior classification performance with substantially higher efficiency, compared to sparse-coding based dictionary algorithms. Sparse coding is an effective building block of learning visual features. A good feature representation is critical for machine learning algorithms to achieve satisfactory results. In recent years, unsupervised feature learning has received increasing research interest in various computer vision and pattern recognition problems. Unlike humanengineered feature extractors that typically require domain knowledge and a large amount of labeled data, unsupervised learning algorithms are generic and designed to automatically discover the intrinsic patterns from the abundant unlabeled data that are usually readily available (from Internet) and require no laborious human labeling. In this dissertation, we will explore the capability of feature learning algorithms in automated biomedical image analysis. Specifically, we will present two unsupervised feature learning models for histopathology image classification. We will also introduce a novel convolutional regression model for nuclei segmentation. Experiments on biomedical image classification and segmentation benchmarks demonstrate that the proposed feature learning systems can achieve very competitive results compared to dedicated systems incorporating biological prior knowledge. Finally, we propose a sparse coding based framework for classifying complicated human gestures represented as multi-variate time series (MTS). Specifically, we will present a novel feature extraction strategy, which can overcome the problem of inconsistent lengths among MTS data and is robust to the large variability within human gestures. Moreover, we will introduce a generic approach to kernelize sparse representation, which leads to enhanced classification performance. Extensive experiments verify the effectiveness of the proposed framework.