Labeling Defective Regions in In-situ Optical Tomography Images

Methods for automatically detecting defects are highly sought after in the world of non-destructive testing and evaluation (NDT&E). Machine Learning (ML) and Deep Learning (DL) algorithms have performed well in this area but often require labeled training examples [1, 2]. This investigation aims to provide insight into the process of obtaining labeled training examples from NDE image data for application to ML and DL. Training data typically consists of the raw NDE data and its corresponding labels. When the NDE data is in the form of an image, the labels can be binary masks, bounding boxes, or semantically segmented images, to name a few. What precisely the labels are labeling depends on the goals in mind. When the goal is the detection of defects, label production might entail binary masking of defective regions, defining bounding boxes around defective regions, or semantically segmenting the image into background, foreground, and defect regions. The performance of a given ML/DL model depends on the quality of the features used for training [3]. The need for accurate defect detection methodologies is particularly stark in metal additive manufacturing (AM) processes, which are prone to producing numerous and disparate process defects. This propensity for metal AM processes to produce defects severely limits their incorporation into end-use component production lines. Methodologies that can accurately detect defects during the manufacturing process are critical to additively manufactured component qualification efforts. This investigation details one step in the ML/DL training process, specifically the defect labeling process, applied to optical tomography images obtained from in-situ monitoring the selective laser melting (SLM) process. The entire image processing workflow entails binary segmentation (masking) of the defective regions, estimation of the defect contours, and estimation of the bounding boxes for defects in the image.