Computer Vision with Active Learning
Machine Vision methods benefit from improving models, tuning trained parameters, or labeling representative data. In a series of experiments, this work validates the hypothesis that Active Learning improves the accuracy of these models. By extending the pseudolabel framework to Active Learning, this work includes a One-shot-learning approach to learn novel image categories by utilising an algorithmic recommender, an online Graphical User Interface to optimise the online Exploration/Exploitation tradeoff for tagging, and a two-step offline binary Active Learning framework to improve the quality of data used for Font Capture. By demonstrating the benefit of Active Learning in these approaches, this work contributes to the hypothesis, as well as concrete Machine Vision applications.
Computer Vision, Object Classification, Semi-supervised Learning, Active Learning, Transfer Learning