The extraction of ancient costume line drawings aims to precisely obtain contour and shape information to aid in re-creation and traditional preservation. However, existing methods increase network depth to improve generalization, leading to a significant increase in the number of model parameters. Therefore, this paper proposes a two-stage edge detection method based on Transformer, aiming to solve the problems of local information loss in images and large model parameter sizes. The first stage divides the image into 16×16 coarse-grained patches and uses an encoder to perform global self-attention calculations to capture dependencies between patches; the second stage covers the image with an 8×8 fine-grained non-overlapping sliding window and calculates the attention within the window through a local encoder to effectively capture subtle edges and reduce costs. A lightweight feature fusion module is designed to support efficient integration of global and local features. Experimental results show that this method outperforms existing methods in extracting edge contour information on ancient costume and public datasets, with an average improvement of 15.9% in the ODS metric. Although OIS and AP does not surpass Informative Drawing, this method shows obvious advantages in model size and time consumption.