In the face of challenges in complex image analysis, particularly for Thangka images with significant cultural value, edge detection encounters the dual difficulties of insufficient accuracy in traditional methods and the dependence on manual annotations in deep learning approaches. To address these issues, this study proposes a self-supervised learning method, constructs a Thangka dataset encompassing four major categories, and designs the ST-PEdger model. This model generates high-quality pseudo-labels through a dynamic-weight multi-threshold Canny edge fusion strategy and simulates real labels for iterative training, thus optimizing the edge detection performance. This approach eliminates the need for manual annotations, avoiding performance degradation caused by annotation discrepancies. Experimental results show that ST-PEdger significantly improves edge detection for Thangka images, reduces the reliance on manual annotations, and holds significant implications for the digital preservation and transmission of cultural heritage.