PSOM: An Vector Primitive Shape Optimization Module

Currently, the most advanced image vectorization approach is based on a learning model, which uses deep learning to predict the parameters of the vector primitives and adjust the shapes of the primitives. These methods mainly focus on the overall accuracy and performance of the vectorization task. However, they do not pay attention to the influence of primitive shape on the quality of generated graphs. The movement of Bezier curve control points will lead to the problem of selfintersection of curves, which makes the vector graphs present a complex topology. In order to solve the above problems, we propose a general primitive shape optimization module, which suppresses the self-intersection of curves and encourages primitives to form simple shapes by controlling the steering between adjacent control points, so as to improve the beauty of primitives and the quality of vector graphics. The experimental results show that the module can significantly improve the visual effect and evaluation indexes of vector images on the two public datasets of Fonts and Emoji.