基于深度学习的印刷纹理图矢量化

Printing Texture Image Vectorization Based on Deep Learning

将纹理扫描图转换成矢量图以供编辑和再利用是计算机图形学和图像处理的一大热点问题。大部分印刷纸面上都包含底纹图案,经过印刷这一复杂物理过程,加上线条本身就相互交错,即使经过高分辨率扫描仪生成位图,线条仍旧会有粘连、边缘信息损失严重、噪声较多的情况,这给纹理线条矢量化带来了很大挑战。采用深度学习的思想,提出一种基于Basnet结构的改进的分割网络,将线条重建转化为图像分割问题来进行图像预处理。该网络能够通过多尺度残差优化将线条模糊的边缘去除,保留以线条骨架为中心的一定宽度的光滑线条。然后使用改进的基于多向量场的线条矢量化算法进行拓扑分析、提取以及矢量化操作。在多种类别扫描图的实验下,验证了该算法能够适用于大部分印刷底纹图案,取得了较好的矢量化效果。

It is a hot issue in computer graphics and image processing to convert texture or line scan images into vector images for editing and reuse.Most of the printed paper contains shading patterns.After the complex physical process of printing,and the lines themselves are interlaced,even though the bitmap is generated by a high-resolution scanner,the lines will still have adhesion,serious edge information loss,and more noise,which brings great challenges to texture line reconstruction and vectorization.In this paper,an improved segmentation network based on Basnet structure is proposed with the idea of deep learning,which transforms line reconstruction into image segmentation for image preprocessing.The network can remove the blurry edges of lines through multi-scale residual optimization,and retain a certain width of smooth lines centered on the line skeleton.Then,the improved line vectorization algorithm based on multi vector fields is used for topology analysis,extraction and vectorization.Finally,this paper also proposes an algorithm for merging and recombining two-dimensional vector data,which can further improve topology and vector data.Through the experiments on various kinds of scanned images,it is verified that the algorithm can be applied to most of the print shading patterns,and has achieved good vectorization effect.