Unsupervised Defect Classification of 2D SEM and 3D X-Ray CT Images from Laser Powder Bed Fusion

This work discusses a method to classify defects in laser powder bed fusion using 2D images of layer samples taken by Scanning Electron Microscope (SEM) and 3D image stacks of a full part by X-Ray Computed Tomography (XCT). Images using SEM are taken of a sampled layer in a printed part and unsupervised classification of defects in the SEM images is performed with Otsu’s thresholding method, K-means classification, and the Robust Automatic Threshold Selection algorithm. The performance of the classifiers, measured against human-generated ground truth defect labels, is improved by registering and fusing multiple SEM images taken under different settings and detector locations. Otsu’s method is shown to be the best classifier for the 3D XCT dataset. Finally, the 2D sample is located in the 3D XCT array and the reliability of the 3D defect classification technique is validated.