非对称内核卷积结合语义置信嵌入的模糊人脸图像重建  

Blurry Face Image Reconstruction Based on Asymmetric Kernel Convolution Combined with Semantic Confidence Embedding

作　　者：胡正平[1,2] 潘佩云 郑赛月 赵梦瑶 毕帅 刘洋 HU Zhengping;PAN Peiyun;ZHENG Saiyue;ZHAO Mengyao;BI Shuai;LIU Yang(School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004;Hebei Key Laboratory of Information Transmission and Signal Processing,Yanshan University,Qinhuangdao 066004;不详)

机构地区：[1]燕山大学信息科学与工程学院,秦皇岛066004 [2]燕山大学河北省信息传输与信号处理重点实验室,秦皇岛066004 [3]不详

出　　处：《模式识别与人工智能》2021年第7期646-654,共9页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金面上项目(No.61771420);河北省自然科学基金项目(No.F2016203422)资助。

摘　　要：受异构卷积原理的启发,在深度学习框架下提出非对称内核卷积结合语义置信嵌入的模糊人脸图像重建网络.针对对称方形卷积内核在进行特征提取时对重要特征表达的不足,使用非对称内核替代,增强方形卷积内核特征的表达能力.在重建阶段,结合非对称内核卷积与语义置信网络,进一步提取每类语义信息在重建中最利于重建效果的特征,结合置信度引导网络向更利于重建的方向训练.在CelebA、Helen数据集上的实验证实文中网络重建效果较优.Inspired by the principle of heterogeneous convolution,a blurry face image reconstruction algorithm based on asymmetric kernel convolution combined with semantic confidence embedding is proposed under the framework of deep learning.Aiming at the deficiency of symmetrical square convolution kernel in expressing important features during feature extraction,asymmetric kernel is employed to replace the symmetric square convolution kernel to enhance the feature expression ability of the square convolution kernel.In the reconstruction stage,the asymmetric kernel convolution is combined with the semantic confidence network to further extract the most benificial features of each type of semantic information for the results in the reconstruction.The confidence is combined to guide the network to train in a more suitable direction for reconstruction.Experimental results on CelebA and Helen datasets show that the proposed algorithm produces better reconstruction results.

关 键 词：图像去模糊 语义网络 卷积神经网络 非对称内核 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]