基于GAN的分步合成人脸素描生成算法

Stepwise Synthetic Face Sketch Generation Algorithm Based on GAN

人脸素描具有丰富的阴影、纹理和鲜明的脸部特征,广泛应用于人脸识别和生活娱乐等领域.鉴于艺术家对人脸素描的绘制步骤具有一定次序的特点,提出一种模拟艺术家绘制人脸素描步骤的算法.将素描生成过程分为2个阶段:第1阶段利用重建的图像辅助生成素描信息;第2阶段利用第1阶段得到的素描信息辅助合成目标素描图像,达到具有脸部轮廓、头发纹理与五官特征、脸部阴影特征的目的.采用CUHK和CUFS数据集进行大量实验,通过PSNR,SSIM和FIR这3个客观评价指标进行对比,结果表明,所提算法的PSNR比典型算法平均提高了4.7 dB,SSIM平均提升0.08,FID分数平均降低8.3;该算法能够生成效果更好的人脸素描图像.

Face sketch has rich shadow,texture and vivid facial features,and is widely used in face recognition and life entertainment.In view of the sequence of the drawing steps of the artist’s face sketch,an algorithm is proposed to simulate the steps of the artist’s face sketch.The process of sketch generation is divided into two stages:the first stage uses the reconstructed image to generate sketch information;In the second stage,the sketch information obtained in the first stage is used to assist the synthesis of the target sketch image to achieve the purpose of having the face outline,hair texture,facial features and facial shadow features.CUHK and CUFS data sets were used to conduct a large number of experiments,and three objective evaluation indicators,PSNR,SSIM and FIR,were compared.The results showed that PSNR of the proposed algorithm increased by 4.7 dB on average,SSIM increased by 0.08 on average,and FID score decreased by 8.3 on average compared with typical algorithms;This algorithm can generate better face sketch images.