Few-shot learning-based generalization detection for challenging part contour in laser powder bed fusion

Abstract

In additive manufacturing process, most existing data-driven methods are trained on high-quality samples with simple geometries from discontinuous layers. Such models, which have extremely poor generalization in challenging samples with small size, complex geometry, bad lighting and ambiguous boundary. This work proposes a Progressive Coarse-to-Fine Network (PCFNet) for fine contour detection of the parts with different geometries. Firstly, This paper provides a mixed part contour dataset comprising a large number of simple samples and a small number of challenging samples. Furthermore, the dilemma of data-driven methods for LPBF online detection is examined by decoupling the similarity between the high-level features of the representative sampled images. Finally, combined with the few-shot learning can improve PCFNet's fine detection level for simple samples and generalization performance for challenging ones. The quantitative and qualitative results on both public and our own datasets demonstrate that the proposed PCFNet exhibits superior detection and generalization performance, significantly outperforming twelve state-of-the-art detection methods. The methodology can detect the layer-wise contour of an orthotropic lattice with the diameter of $0.25mm$ (≈3 pixels), achieving a statistical fidelity exceeding 84% (mIoU).