Advancing image processing through vector graphics techniques

Unlike discrete raster images composed of pixels, vector images represent content using continuous geometric primitives, offering advantages such as resolution-independence and easier editing. Common primitives include curves and meshes, and different primitives yield different effects. In his thesis, Kuanhao He presents contributions in three directions.

First, He addresses mesh-based image vectorization. Existing methods often produce dense meshes for images with high-frequency details. He proposes a novel method that automatically vectorizes images into sparse quadrilateral Coons patches. Patch layouts are guided by a harmonic cross field constrained by image features, supporting T-junctions. Mesh colors further enhance adaptability, enabling local adjustment to feature density. Experiments demonstrate the method’s utility, accuracy, and sparsity.

Second, He focuses on vector graphics optimization. Even good vectorizations often fall short in geometry and color. He developed a method that optimizes curved triangle meshes using a differentiable mapping between mesh parameters and the difference from the input image. Treating the image as a bilinear spline and using Monte Carlo integration allows parallelizable operations. Tests show effective and efficient quality improvement.

Finally, He tackled image-based sketching. His algorithm transforms images into pencil-stroke illustrations conveying geometry. By integrating learning-based depth estimation, He generates a frame field to guide stroke directions, capturing object shapes accurately. Experiments demonstrate effective geometric detail preservation.

Dissertation: https://hdl.handle.net/11370/97725946-2ecb-4297-b213-6050a77df204