Various areas produce huge high-dimensional data streams which need to be monitored to detect parts of or locations in the data stream indicating faults or critical events. We consider the specific problem to analyze image data, in order to detect automatically suspicious areas. For this purpose the image is locally smoothed to reduce noise and justify Gaussian approximations. These statistics are standardized by an estimate based on residuals calculated from a smooth estimate of the in-control background which is allowed to vary but on a larger scale. Based on a new result related to Berman's inequality for maxima in Gaussian series, a procedure is proposed which uses Gumbel-type thresholds for detection. Simulations show that for a properly chosen bandwidth of the local statistics, the approximation of the significance level is accurate for realistic image resolutions. The method is illustrated by an application to electroluminescence images of solar modules.
Part of this work has been supported by a grant from German Federal Ministry of Economics and Energy (BMWi), #0325588B.
|