Learning with limited annotations: Deep semi-supervised learning paradigm for layer-wise defect detection in laser powder bed fusion

Powder bed quality is critical to the quality of parts manufactured by laser powder bed fusion (LPBF) and mass production. Recently, numerous powder bed defect detection methods based on semantic segmentation algorithms have been developed. However, these data-driven approaches face the indispensable challenge of insufficient annotated data. Especially in the context of defect segmentation in Additive Manufacturing (AM), pixel-wise image labeling is time-consuming and demands substantial prior knowledge. Semi-supervised Learning (SSL) can leverage unlabeled data to enhance the training process of deep learning models. During the layer-by-layer forming process in LPBF, thousands of powder bed images can be obtained but most of them remain unused because of the lack of annotations, which fully satisfy the situation with semi-supervised learning. To address the above issue, this paper proposes a deep semi-supervised learning-based paradigm for powder bed defect segmentation, allowing model learning with limited annotated data. Concretely, the proposed paradigm generates pseudo-labels for unlabeled data, enabling the utilization of a substantial amount of unlabeled data in the manufacturing process. Aiming at the issue of low-quality pseudo-labels generated from low-quality unlabeled data, we employ Mean Teacher Framework to separate the generation of pseudo-labels. Moreover, aiming at the lack of data diversity, we employ Consistency Regularization to enhance the model's generalization performance. Additionally, we created a dataset comprising 406 images of powder-bed defects, with each image annotated at the pixel level. Extensive experiments on the dataset have shown the proposed paradigm's effectiveness over supervised methods, even with limited labeled data (only 1/8 annotated).