摘要
以往修复图像的办法是将任意缺失区域的推断应用到损失的部分中,难以得到高精度的复原图像.而基于生成式对抗网络(GAN)结合二次优化算法可以对图像损失部分进行修复.该框架对生成式对抗网络结构进行了改进,选用残差神经网络结构替换生成器结构,可生成更有效的伪造图像集,从而激发判别器提升其性能,并选用优先级函数和均方误差(MSE)确定待修复补丁和与其最佳匹配补丁,结合期望最大化(EM)算法最小化来优化补丁匹配与补丁合成的细节,提高图像复原的准确度。
Due to the conventional solution of image completion,which the estimation of an arbitrary missing area is pasted into the lost portion,it is difficult to obtain a high-precision restored image. This paper proposed an image restoration based on the generative adversarial networks( GAN) combined with the quadratic optimization algorithm. The framework-that by improving the structure of GAN and replacing the generator structure with the residual neural network structure-generates a more effective pseudo image set,which stimulates the discriminator to improve its performance. Meanwhile,the framework improves the accuracy of image restoration by selecting the priority function and the mean square error( MSE) to determine which patch to be repaired and its best matching patch,and in combination with minimizing the expectation maximization( EM) algorithm to optimize the details of patch matching and patch synthesis.
作者
潘玥
杨会成
储慧敏
PAN Yue;YANG Huicheng;CHU Huimin(School of Electrical Engineering,Anhui Polytechnic University,Wuhu Anhui 241000,China)
出处
《海南热带海洋学院学报》
2020年第2期81-87,共7页
Journal of Hainan Tropical Ocean University
基金
安徽省高校自然科学研究重点项目(KJ2018A0122)。
关键词
生成式对抗网络
残差网络
优先级函数
均方误差
期望最大化算法
generative adversarial networks
residual network
priority function
mean square error
expectation maximization algorithm
