Philipp Klein (Otto von Guericke University Magdeburg)

​

Title: Anomaly detection based on MOSUM statistics in large image data

 

Abstract - Being able to identify anomalies in image data is important in many practical applications. This includes the detection of cancerous cells in medicine, obstacles in autonomous driving and fissures in the construction material of buildings. Oftentimes it is helpful to have an efficient pre-processing method in order to identify regions potentially containing anomalies, especially when these anomalies are sparse in large data.

 

In this talk, we present a procedure for the detection of fissures in 2D image data based on MOSUM statistics. In particular, based on the observation of local geometric properties of fissures and natural anomalies in concrete, we use a combination of moving semicircle- and rectangle-shaped windows in order to identify regions that potentially have fissures. Furthermore, under the null hypothesis of no anomalies for image data in dimension P, we are able to show convergence for MOSUM statistics using a general class of scan windows towards a functional of P-parameter Gaussian processes. In particular, this class includes 2D convex sets, which we use for fissure detection.

​

​

​

RETURN TO SEMINAR HOMEPAGE