Machine Learning for Large Image Datasets
Oral Defence Date:
Professors Kazunori Okada, Dragutin Petkovic, and William Hsu
This thesis looks at machine learning in two contexts. First, we apply CBIR to medical image analysis. While previous studies focused on feature design, our study focuses on metric design. Our technique learns a metric using information theoretic metric learning. We compare our learned metric on a SIFT bag-of-words system against a priori measures from literature and evaluate it using the ImageCLEF-2011 benchmark. Our results show the advantage of this metric learning approach and of L1-distance based measures. The second study is motivated by the need of X-ray crystallograhers to elucidate molecular structure from quality crystallized proteins, and by the scarcity of successful crystals. The process is human-curated, time-consuming, and error-prone and we aim to automate search and classification using machine learning. Our results prefer elastic net and cascade classifiers composing random forest with linear discriminant or naive Bayesian techniques.
machine learning, image processing, computer vision, statistics, random forest, elastic net, classification, image retrieval, CBIR, metric, metric learning, X-ray crystallography, protein crystallization, feature selection, model selection, feature design, metric design, big data, cascade classifier, imbalanced datasets, bag of words, sift, K-means clustering, ITML, Gabor wavelet, GLCM, Shape context