摘要
拍摄人像时,由于摄影器材震动或人物移动等原因,图像会出现运动模糊情况,严重影响了画面品质。而在影像跟踪或门禁识别等场景中,若画面出现运动模糊,可能导致目标无法识别,使得定位、辨认、追踪等任务失败。因此,去除运动模糊对于人脸识别应用有着至关重要的作用。文章针对人脸图像的运动模糊问题,提出一种基于生成对抗网络的人脸运动模糊去除方法,在编解码结构中引入多个跳跃连接,将卷积过程中提取的特征引入反卷积过程,同时通过全局跳跃连接,提高对特征信息的复用,并降低学习复杂度,最后调整损失函数权重,得到模糊图像到修复图像的端到端网络。实验结果表明,该方法在消除人脸运动模糊方面有较好的效果,对人脸的轮廓等细节的恢复也有很好的改善作用。
When taking portraits,due to the vibration of photography equipment or the movement of people,the image may experience motion blur,which seriously affects the image quality.In scenarios such as image tracking or access control recognition,if there is motion blur on the screen,it may lead to the inability to recognize the target,resulting in the failure of tasks such as localization,recognition and tracking.Therefore,removing motion blur plays a crucial role in facial recognition applications.In this paper,we propose a method based on generation countermeasure network for motion blur of face images,which introduces multiple jump connections into the codec structure,introduces the features extracted in the convolution process into the deconvolution process,improves the reuse of feature information through global jump connections,reduces the learning complexity,and finally adjusts the weight of the loss function to obtain the end-to-end network from the blurred image to the repaired image.The experimental results show that this method has a good effect in eliminating facial motion blur and also improves the restoration of details such as facial contours.
作者
曾孟佳
户哲
黄旭
Zeng Mengjia;Hu Zhe;Huang Xu(School of Information Engineering,Huzhou University,Huzhou 313000,China;School of Electronic Information,Huzhou College,Huzhou 313000,China;Huzhou Key Laboratory of Urban Multidimensional Perception and Intelligent Computing,Huzhou 313000,China)
出处
《无线互联科技》
2024年第8期66-70,共5页
Wireless Internet Technology
基金
教育部人文社会科学一般项目,项目编号:20YJCZH005
浙江省湖州市工业攻关项目,项目编号:2018GG29
湖州学院国家级大学生创新创业训练项目,项目编号:202313287007。
关键词
人脸图像
运动模糊
生成对抗网络
损失函数
facial images
motion blur
generating adversarial networks
loss function
