Stitch-Based Image Stylization for Thread Art Using Sparse Modeling

Abstract

Random-needle Embroidery (RNE) is a graceful Chinese Embroidery art enrolled in the World Intangible Heritage. In this paper, we propose a rendering method to translate a reference image into an art image with the style of random-needles. Since RNE artists create artwork by stitching thousands of intersecting threads with complex patterns into an embroidery cloth, the key of RNE rendering is to define its threads distributions in vector space (actual physical space) and generate its artistic styles in pixel space (coordinate system of the image). To this end, we first define “stitch” which is a collection of threads arranged in a certain pattern as the basic rendering primitive. A vector space stitch model is presented, which can automatically generate various thread distributions in stitches. Then, the rendering primitives are generated by rasterizing the stitches on 2D pixel arrays. During runtime, new stitches can be synthesized to portray the image content via sparse modeling based on the pre-defined stitches. In order to avoid mosaic effects, this result is further refined by incorporating local stitch vector constraints, in which we enforce the thread distribution of the local stitch to be similar to its adjacent stitches. Finally, rendering image is generated by placing stitches with different attributes on the canvas. Experiments show that our method can perform fine images with the style of random-needle.