Learning sparse features and metric in signal and image processing

Feature learning;Sparse coding;Metric learning;Dictionary learning;Spectral analysis;Image classification;Wildlife monitoring;Shape matching

Wang, Jiangping

关键词: Feature learning;    Sparse coding;    Metric learning;    Dictionary learning;    Spectral analysis;    Image classification;    Wildlife monitoring;    Shape matching;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/98199/WANG-DISSERTATION-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This dissertation studies two aspects of feature learning: representation learning and metric in feature space, from a machine learning perspective. Feature learning is a fundamental problem in computer vision and machine learning. First introduced in computational neuroscience in the context of sparse coding in the visual system, sparse coding plays a key role in feature representation learning, as the over-complete dictionary allows more representation flexibility and efficiency, and captures structures and patterns inherent in the raw data. First, we explore sparse representation and propose a novel sparse matrix factorization method for learning a dictionary in both a reconstructive and discriminative manner. The obtained representations can be directly used for multi-class classification. We also apply sparse representation on a camera trap dataset for wildlife monitoring and demonstrate good performance of sparse features in a challenging real world scenario. Second, from a theoretical perspective, we show that the popular ℓ1-norm based methods break down in the presence of high coherence and large noise. We introduce a novel regularization approach to handle model collinearity and obtain parsimonious variable selection simultaneously. The regularization term is non-convex and can take into account structured sparsity (e.g., group sparsity). We propose an efficient iterative thresholding procedure for solving the optimization. Our method achieves state-of-the-art performance for super-resolution signal spectrum estimation.On the other hand, in order to enhance the discriminative power of learned features, supervised learning is crucial. A proper metric is desired for various problems, such as image retrieval, similarity learning and face verification. We propose a ranking based metric learning algorithm under maximum margin criterion. We propose both batch and online algorithms. The regret bounds are given for online algorithms. Experiments are conducted on 3D human body shape matching problem and state-of-the-art performance is achieved.