一种强真实感的图像生成算法设计与仿真  

Design and Simulation of Highly Realistic Image Generation Algorithm

作　　者：应卫强[1,2] 张帆 张玲燕[2] YING Wei-qiang;ZHANG Fan;ZHANG Ling-yan(City College,Zhejiang University,Hangzhou Zhejiang 310015,China;School of Software,Zhejiang University,Hangzhou Zhejiang 310027,China)

机构地区：[1]浙江大学城市学院,浙江杭州310015 [2]浙江大学软件学院,浙江杭州310027

出　　处：《计算机仿真》2022年第4期492-495,500,共5页Computer Simulation

摘　　要：传统方法下生成的跨模态图像易造成目标部分重要信息缺失,生成的图像缺乏真实感,于是提出改进生成式对抗网络和半监督学习的跨模态图像生成方法。建立生成式对抗网络,分析半监督学习特征,经融合后组成半监督生成式对抗网络。在判别器中使用卷积神经网络、在生成器中引入反卷积神经网络,在半监督生成式对抗网络中添加分类器,改进所建的网络模型,利用全变差正则化项建立伪判别损失函数。利用架构的随机梯度下降优化算法,完成散度似然比的直接优化,最后在三个网络的共同作用下,输出生成的跨模态图像。仿真阶段分别从视觉效果与评估指标两个角度,验证出所提方法的有效性,结果证明上述方法不仅使生成图像更具真实感,而且保留了大部分目标特征,能够满足图像的高质量需求。In order to ameliorate the realism of cross modal images generated by traditional methods, this paper reports a cross modal image generation method based on improved generation anti network and semi-supervised learning. The generative countermeasure network was established and the semi-supervised learning characteristics were investigated to form a semi-supervised generative countermeasure network. Convolutional neural networks were used in discriminators. In the generator, deconvolution neural network was introduced. In the semi-supervised generative countermeasure network, a classifier was added to improve the network model. The total variation was applied to the de-regularization term to establish the pseudo discriminant loss function. The random gradient descent optimization algorithm of the architecture was used to optimize the divergence likelihood ratio. Eventually, under the joint action of the three networks, the generated cross modal image was output. The simulation experiments verified the effectiveness of the proposed method. The results show that the image generated by this method is more realistic, retains most of the target features, and meets the high-quality requirements of the image.

关 键 词：生成式对抗网络 半监督学习 跨模态图像生成 随机梯度下降优化算法 损失函数 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]