摘要
随着各国经济贸易、文化交流往来的日益频繁,快速、有效地区分各国人员身份是当前人脸识别领域的一项重要研究。本文特针对亚洲区域五个国家(中国、日本、韩国、泰国、印度)进行人脸分类识别的研究,本文基于MobileNet进行五国人脸分类识别,因为这五国人脸较为相似,为能有效降低冗余,本文将八度卷积插入该网络中减少冗余,提升精度;并提出使用中心损失函数和交叉熵损失函数相结合的方法来提升准确率。经过实验验证,本文提出的在网络中插入八度卷积和中心损失函数两种改进方法均可以提升准确率,其最高准确率可达87.84%,其Error top 1最低达到0.120%。
With the increasingly frequent economic, trade and cultural exchanges between countries, quickly and effectively distinguishing the identity of people in various countries is an important research in the field of face recognition. This paper focuses on the research of face classification and recognition in five Asian countries (Chinese, Japanese, Korean, Thailand and Indian). In this paper, face classification and recognition in five Asian countries are based on MobileNet. Because the faces of these five countries are similar, to reduce redundancy in this paper, octave convolution is inserted into the network to reduce redundancy and improve accuracy;and a method using a combination of center loss function and cross-entropy loss function is proposed to improve accuracy. After experimental verification, both the octave convolution and the center loss function proposed in this paper can improve the accuracy rate, and the highest accuracy rate can reach 87.84%, its Error top 1 is at least 0.120%.
出处
《图像与信号处理》
2020年第3期146-155,共10页
Journal of Image and Signal Processing
关键词
多国人脸分类
八度卷积
中心损失函数
宽度乘子
倒残差模块
Multi-National Face Classification
Octave Convolution
Center Loss
Width Multiplier
Inverted Residuals
