Our paper “Visual inspection via anomaly detection by automated uncertainty propagation” got accepted for this year’s SPIE Photonics Europe meeting in Strasbourg and will be presented in the poster session on April 5.
The core idea of the paper is to capture the intended appearance of test objects in a visual inspection scenario by calculating pixel-wise statistics (mean and standard deviation) for a set of defect-free training objects. During the inspection, acquired inspection images are tested for outliers (w.r.t. the calculated statistics of the defect-free objects). Such outliers could represent material defects. Since any image processing operations invalidate the calculated statistics, we show that using Gaussian uncertainty propagation in concert with automatic differentiation allows to automaically and conveniently update the statistics with respect to the performed image processing operations.