Semantic Segmentation of Selective Laser Melting Powder Bed Images Based on Multi-Scale Feature Fusion and Hierarchical Transfer Learning

Abstract

Layer-wise powder bed images acquired during selective laser melting (SLM) can characterize the surface conditions after powder spreading and laser scanning, thereby providing essential information for in situ quality monitoring. Semantic segmentation enables pixel-level identification and localization of powder bed defects. However, accurate segmentation remains challenging because different defect types exhibit substantial variations in morphology and scale, while class imbalance and cross-device imaging domain shifts further complicate the task. To address the challenges of multiple defect categories, significant scale variations, imbalanced class distributions, and imaging domain shifts across different devices in powder bed image monitoring for selective laser melting (SLM), a semantic segmentation method based on multi-scale feature fusion and hierarchical transfer learning is proposed. First, post-spreading and post-scanning images from the same layer are combined to form a dual-channel input. Four parallel branches are then designed to extract global features, native-resolution details, local morphological features, and positional encodings, respectively, enabling multi-scale feature fusion to capture defect information across different spatial scales. Second, a two-stage transfer learning strategy is developed for cross-device applications. In the first stage, network parameters are updated differentially according to the domain sensitivity of individual modules. In the second stage, the entire network is jointly optimized using a small learning rate, while a maximum mean discrepancy (MMD) constraint is introduced to align the feature distributions of the source- and target-domains. Experimental results show that the proposed method achieves a mean Intersection over Union (mIoU) of 0.879. After hierarchical transfer learning, an mIoU of 0.855 is obtained on the experimental dataset from the target-domain. These results demonstrate that the proposed method effectively improves the segmentation accuracy of complex powder bed defects and exhibits strong adaptability across different devices.

The authors' abstract, as published at the source. Electronics, 2026 · DOI ↗