Human figures on Chinese export porcelain are a vital resource for art-historical study, yet their analysis is hindered by the limitations of manual, subjective methods. The precise segmentation of these figures from intricate backgrounds is a significant technical challenge, complicated by stylistic diversity and ambiguous object boundaries. To automate and scale this process, we propose Porcelain-SegNet, a novel, boundary-aware deep learning framework designed specifically for this task. We first introduce EPF-Seg-1k, the first large-scale benchmark dataset for segmenting figures on porcelain, comprising over 1,000 high-resolution images with meticulous, pixel-level annotations. Our Porcelain-SegNet architecture features a Swin Transformer backbone to effectively model global context, attention-guided skip connections (CBAM) to enhance feature fusion by suppressing background noise, and a composite, boundary-aware loss function to ensure the generation of crisp and accurate outlines. Extensive experiments and ablation studies demonstrate that our method achieves state-of-the-art performance on the EPF-Seg-1k test set, obtaining a mean Intersection over Union (mIoU) of 90.4% and significantly outperforming established segmentation models. The high-fidelity masks produced by our model provide a robust computational foundation for enabling new forms of large-scale, quantitative analysis of iconographic and stylistic features in the field of digital art history.