GPU加速的近实时图像彩色化  

GPU-Accelerated Near Real-Time Image Colorization

作　　者：赵汉理[1] 季智坚 金小刚[2] 厉旭杰[1] Zhao Hanli;Ji Zhijian;Jin Xiaogang;Li Xujie(Intelligent Information Systems Institute, Wenzhou University, Wenzhou 325035;State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou 310058)

机构地区：[1]温州大学智能信息系统研究所,温州325035 [2]浙江大学CAD&CG国家重点实验室,杭州310058

出　　处：《计算机辅助设计与图形学学报》2017年第8期1425-1433,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61472351);浙江省自然科学基金(LY15F020019;LQ14F020006);温州市科技计划项目(G20150019)

摘　　要：灰度图像彩色化技术需要人工交互来完成彩色化优化过程,针对现有的方法只注重彩色化的效果而忽略算法的执行效率,严重影响了用户的交互体验的问题,充分利用GPU的高性能并行优势,提出基于GPU加速的近实时图像彩色化方法.在预处理阶段,运用基于图像块纹理特征的Patch Match算法在全局图像空间高效地查找每个像素的K最近邻,并提出基于压缩表示的对称稀疏矩阵并行构造算法来保证着色线条的颜色在图像近邻像素之间的对等传播;在用户交互阶段,根据用户输入构建能量函数,并运用并行共轭梯度法计算出彩色图像的颜色.实验结果表明,该方法不但能生成高质量的图像彩色化效果,而且图像彩色化过程具有近实时性的处理性能.Grayscale image colorization requires user interactions to complete the colorization optimization process.Existing methods only aims to the colorization effect and do not take into consideration the interaction experience.A novel GPU-accelerated near real-time image colorization method is proposed by taking advantage ofhigh-performance parallelism of the GPU.In the preprocessing step,this method uses a patch-based PatchMatchalgorithm to search K-nearest neighbors for each pixel on the global image space and then introduces a parallelconstruction algorithm for the compressed symmetry sparse matrix to ensure the equivalent propagation of colorsbetween neighboring pixels.During the user interaction,this method builds an energy function according to theuser input and outputs the colorized image by using the parallel conjugate gradient solver.Experimental resultsshow that the proposed method can not only produce high-quality colorized images but also achieve near real-time performance.

关 键 词：图像彩色化 K最近邻 压缩稀疏矩阵 图形处理单元 

分 类 号：TP317.41[自动化与计算机技术—计算机软件与理论]