摘要
人脸识别已经广泛地应用于日常生活中,作为关键技术之一的人脸清晰度评价成为了热门的研究课题.然而,传统的手工提取特征的方法在效果和鲁棒性上都有所欠缺.为此,我们运用卷积神经网络实现特征的构造和选择,有助于提高评价结果的准确率.同时针对网络复杂、参数过多和耗时长等问题,还提出将传统的卷积结构改造成双卷积层结构的方法来提升计算速度.经过大量的实验表明,本文提出的人脸清晰度评价算法能够准确地进行人脸清晰度的评估,并且具有较快的处理速度.
Face recognition technology has been widely used in daily life.As one of the key technologies,face sharpness evaluation got much attention.However,there were less effectiveness and robustness in using traditional methods with manual features designed.We studied a method to construct and select features by convolutional neural networks,which could improve the accuracy of face recognition. Meanwhile,the structure of double convolution layers was proposed to solve the problems such as complicated network calculation,too many parameters and much calculation time consumed.Experiments demonstrated that the face sharpness evaluation algorithm had a better accuracy and a fast processing speed.
出处
《中国计量大学学报》
2017年第4期509-515,共7页
Journal of China University of Metrology
基金
浙江省自然科学基金资助项目(No.LY15F020021)
浙江省科技厅公益性项目(No.2016C31079)
关键词
深度学习
清晰度评价
图像分类
视频监控
deep learning
sharpness assessment
image classification
video surveillance