Few-shot interactive segmentation network for assisted annotation of coated particle images

Accurate measurement of the coating layer thickness in high-temperature gas-cooled reactor (HTGR) fuel elements is a critical technical requirement for reactor safety assessment. Our previous work achieved automatic coating thickness measurement by integrating ceramographic method with machine vision. However, when adapting to newly designed particle configurations, the deep learning models suffer from low responsiveness due to the heavy dependence on large volumes of annotated training data. Although interactive image segmentation shows promise for addressing this issue, existing methods generally rely on large-scale datasets, limiting their applicability in few-shot industrial scenarios. To tackle this challenge, we propose an interactive segmentation framework tailored for coated particle images. The primary technical contributions of this study are threefold: (1) A lightweight Vision Transformer architecture is constructed with optimized parameters, enabling end-to-end training while maintaining robust feature encoding performance; (2) An Attentional Semantics & Detail Fusion module is designed to enhance multi-scale feature extraction and mitigate detail loss caused by backbone compression; (3) A Self-Attention Segmentation Head is developed to improve the representation of key regions by leveraging self-attention mechanisms. Experimental results on our coated particle dataset demonstrate that the proposed method achieves 95 % Intersection over Union (IoU) with an average of only 2.5 user clicks. When the number of interactions increases to five, the mean IoU (mIoU) reaches 97.12 %, consistently outperforming existing interactive segmentation networks.