摘要
现有大部分基于卷积神经网络的图像去模糊算法,在人脸识别相关场景应用中,存在识别率提升不明显、参数量过多等问题.针对上述现象,本文提出一种基于条件生成对抗网络(Conditional Generative Adversarial Networks,DGAN)的图像去模糊算法.该算法将结合了轻量级分组卷积与改进SE(Squeeze-and-Excitation,SE)注意力机制的Group-SE模块作为生成器的主体部件,将引入了全局性稠密连接的改进DenseNet作为判别器核心,以解决去模糊技术应用在人脸识别算法中的低效率等问题.在CASIA WebFace和LFW数据集上的实验表明,所提算法在提升图像质量、降低模型参数量等方面均有不错的表现,并且采用了该算法作为图像预处理过程的SphereFace、CosFace和ArcFace人脸识别方法,在LFW数据集上识别率分别有4.22%、3.43%和3.95%的大幅提升.
Most of the existing image deblurring algorithms based on convolutional neural networks have problems such as insignificant improvement in recognition rate and too many parameters in face recognition-related scene applications. In response to the above phenomenon,this paper proposes an image deblurring algorithm based on Conditional Generative Adversarial Networks( DGAN). The algorithm uses the Group-SE module that combines lightweight group convolution and improved SE( Squeeze-and-Excitation,SE) attention mechanism as the main part of the generator,and the improved DenseNet with global dense connection is used as the core of discriminator to solve the problem of low efficiency in the application of deblurring technology in face recognition algorithms. Experiments on CASIA WebFace and LFW datasets show that the proposed algorithm has good performance in improving image quality and reducing the amount of model parameters,and uses this algorithm as the image preprocessing process of SphereFace,CosFace and ArcFace face Recognition method,the recognition rate on the LFW data set has been greatly improved by. 22%,. 43% and. 95% respectively.
作者
曾凡智
邹磊
周燕
邱腾达
ZENG Fan-zhi;ZOU Lei;ZHOU Yan;QIU Teng-da(Department of Computer Science,Foshan University,Foshan 528000,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第12期2607-2613,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61972091,61602116)资助
广东省自然科学基金项目(2017A030313388)资助
广东省工程技术研究中心项目(G601624)资助
广东省普通高校重点项目(2019KZDXM007)资助
广东省普通高校科研项目(2020ZDZX3049)资助
佛山市工程技术研究中心项目(2017GA00015)资助
佛山市教育局特色创新项目(2019XJZZ10)资助。
关键词
条件生成对抗网络
图像去模糊
人脸识别
分组卷积
注意力机制
conditional generative adversarial networks
image deblurring
face recognition
group convolution
attention mechanism
