摘要
针对目前人脸到素描合成存在生成的素描图轮廓模糊、细节纹理缺失等问题,提出一种采用循环生成对抗网络(CycleGAN:Cycle-Generative Adversarial Networks)解决方案。构建多尺度CycleGAN,生成器采用深度监督的U-Net++结构为基础,在其解码器端进行下采样密集跳跃连接;在其生成器的编码器端设计通道注意力和和空间注意力机制形成特征增强模块;最后在生成器中增加像素注意力模块。实验结果表明,与现有经典算法相比,从主观视觉评测和利用现有的4种图像质量评价算法进行质量评估,该方法较好地合成了素描图像的几何边缘和面部细节信息,提高了素描图像的质量。
At present,Face sketch synthesis has a series of problems,such as generateing fuzzy outline,lacking of detail texture and so on.Therefore,using CycleGAN(Cycle-Generative Adversarial Networks)as a solution to build multi-scale cyclegan is proposed.Method innovation is mainly reflected in:The generator adopts the deep supervised U-net++structure as the basis,and performs down sampling dense jump connection at its decoder;The encoder end of the generator designs the channel attention and spatial attention mechanism to form a feature enhancement module;a pixel attention module is added to the generator.Compared with some existing classical algorithms,from the subjective visual evaluation and using the existing four image quality evaluation algorithms for quality evaluation,the experimental results show that this algorithm can better synthesize the geometric edge and facial detail information of sketch image,and improve the quality of sketch image.
作者
葛延良
孙笑笑
王冬梅
王肖肖
谭爽
GE Yanliang;SUN Xiaoxiao;WANG Dongmei;WANG Xiaoxiao;TAN Shuang(School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China)
出处
《吉林大学学报（信息科学版）》
CAS
2023年第1期76-83,共8页
Journal of Jilin University（Information Science Edition）
基金
黑龙江省自然科学基金资助项目(LH2020F005)。
关键词
深度学习
多尺度CycleGAN
卷积神经网络
特征增强模块
像素注意力模块
deep learning
multi-scale cycle-generative adversarial networks(CycleGAN)
convolutional neural networks(CNN)
feature enhancement module
pixel attention module
