Adversarial Domain Adaptation for Improved Part-to-Part Generalization of Deep Learning Segmentation Models in Aerosol Jet Printing

Abstract The segmentation performance of deep learning (DL) models is highly dependent on the statistical/probabilistic distribution and characteristics of the data used during their training process. Unfortunately, the increased personalization of additively manufactured products leads manufacturers to use different printing techniques, materials, and parameters. These variations alter the properties of the data collected, causing a statistical domain shift that hinders the generalization of DL models when applied across different parts. This issue is exacerbated when the DL model is supervised and thus requires annotations, a task that is sometimes performed manually and is very time-consuming and tedious. To alleviate this problem, this study proposes the use of adversarial domain adaptation, an unsupervised learning approach that can improve models' generalization without the need for extra annotation. A case study analysis was performed on real microscopic images taken from aerosol-jet-printed samples. The proposed method achieved a 16% improvement in segmenting images across parts printed on two different material substrates.