THESIS
2017
xiii, 39 pages : illustrations ; 30 cm
Abstract
While in surface defects detection applications, detectors based on supervised learning can achieve high accuracy, their requirement of plentiful well-balanced labeled training set and vulnerability to defects which are absent from
training set usually limit their utilizations. Anomaly detectors can be trained
solely on non-defect data by adopting the unsupervised learning approaches,
however, many of existing methods are either infeasible to high dimension scenarios or lack of solid interpretations. In this thesis work, we propose a stochastic anomaly detector based on distribution estimation for spatial data, and
provide an alternative deterministic approach to simplify the detection procedure afterwards. The results from benchmark tests show their proficiency as
anomaly detecto...[
Read more ]
While in surface defects detection applications, detectors based on supervised learning can achieve high accuracy, their requirement of plentiful well-balanced labeled training set and vulnerability to defects which are absent from
training set usually limit their utilizations. Anomaly detectors can be trained
solely on non-defect data by adopting the unsupervised learning approaches,
however, many of existing methods are either infeasible to high dimension scenarios or lack of solid interpretations. In this thesis work, we propose a stochastic anomaly detector based on distribution estimation for spatial data, and
provide an alternative deterministic approach to simplify the detection procedure afterwards. The results from benchmark tests show their proficiency as
anomaly detectors and capability as defect detectors.
Post a Comment