Robust calibration marker detection in powder bed images from laser beam melting processes

Laser beam melting (LBM) systems produce parts by melting metal powder according to the sliced 3D geometry using a laser. After each layer, new powder is deposited and the process is repeated. Process monitoring via acquisition and analysis of layer images during the build job is a promising approach to thorough quality control for LBM. Image analysis requires orthographic images, which are usually not available as the camera cannot be placed directly above the build layer due to the position of the laser window. The resulting perspective distortions have to be corrected before analysis. To this end we compute a homography from four circular markers which are "drawn" into the powder bed by the machine's laser and detected in the acquired images. In this work we present a robust method for the automatic detection of calibration markers, which deals with the noise-like powder regions, disconnected lines, visible support structures and blurred image regions. Our homography estimation method minimizes the shape error between transformed circular reference marker shapes and detected elliptical markers yielding an image with correct aspect ratio and minimal distortions. Our method achieves a detection rate of 96.3 % and a spatial detection error of 2.0 pixels (median, 95 %-percentile: 5.17pixels) compared to a manually created ground truth.