Learning-based Vector Graphics Processing and Synthesis

Abstract

In contrast to the significant developments in pixel-space image understanding and generation, vector graphics remain challenging for recent studies. Although rasterization—the process of converting vector graphics into raster images—is relatively straightforward, vectorization from images poses notable challenges due to its inherent challenges. From the recent studies targeting tasks related to vector graphics, we observe that they often face significant hurdles due to the complexity of representations, the scarcity of data, and inherent difficulties associated with the methodologies employed. This thesis aims to provide a comprehensive understanding of the current state of research in related fields, thereby identifying directions for proposing novel learning-based approaches and improving the performance of existing methods for processing and generating different vector representations.