<p>In recent years, semantic segmentation has achieved significant performance improvements in many application fields. However, due to the fact that elements in Thangka images often exhibit similar texture features and rich yet highly analogous color distributions, the performance of existing algorithms on Thangka datasets remains suboptimal. To address this challenge, we propose a multi-scale edge-guided and detail-enhanced Thangka image segmentation algorithm. Firstly, through the independent training of the detail enhancement module, the detail features are enhanced and the influence of noise on training is reduced. In addition, the window multi-head self-attention pooling pyramid is used to capture the missing global context information, so as to effectively extract high-level semantic information and spatial information. Finally, multi-scale edge guidance preserves the semantic features in each scale feature layer, promotes the guidance and fusion of multi-scale spatial information, and preserves the edge and texture features in spatial information to the greatest extent. We conducted qualitative and quantitative experiments on our self-built high-definition Thangka dataset, achieving improvements of 1.22%, 0.66%,1.04%, 0.75% and 0.46% over state-of-the-art algorithms in terms of mIoU, mPA, Precision, mRecall and Dice metrics, respectively. Additionally, the algorithm’s generalization was validated on the PASCAL VOC2012 dataset.</p>