Image Extraction of Thangka Line Drawings with Transformer

In the drawing process of Thangka, the painter can draw different types of Thangka on the same line drawing, but the painter needs to redraw an identical line drawing every time. The drawing and coloring of the line draft are time-consuming and laborious. Therefore, in view of the difficulty in obtaining the real line drawing image data of Thangka and the distortion of the effect of the existing line drawing extraction methods, this paper proposes a Thangka line drawing extraction method based on Transformer: ETLTER. By introducing Vision Transformer, ETLTER captures coarse-grained global context, medium-grained local context, and fine-grained detail context features simultaneously in the three stages. In addition, the feature fusion module (FFM) fuses the feature information extracted from the three stages to predict the final Thangka manuscript effect. Through the processing results of the above three stages, ETLTER can generate clear and concise Thangka line drawings. Based on our own Thangka image dataset TK1500, the Thangka line drawings extracted by the model in this paper have less noise, and clear lines and are close to the line drawings drawn by the real Thangka painter compared with the existing line drawings extraction methods. The average rank of manuscript images extracted by this method is 1.167, ranking first among the 30 methods. The comprehensive evaluation results show that our methods achieve the state-of-art performance in Thangka line drawing extraction, and ETLTER is of great significance to the training of new Thangka painters.