基于大规模集成学习的正脸姿态估计算法  

Frontal pose assessment algorithm based on large-scale ensemble learning

作　　者：陈婉琪 邓春华 CHEN Wan-qi;DENG Chun-hua(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Institute of Big Data Science and Engineering,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan University of Science and Technology,Wuhan 430065,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]武汉科技大学大数据科学与工程研究院,湖北武汉430065 [3]武汉科技大学湖北省智能信息处理与实时工业系统重点实验室,湖北武汉430065

出　　处：《计算机工程与设计》2024年第12期3592-3599,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61806150)。

摘　　要：现有的正脸姿态估计任务依赖于头部3D姿态的估计和经验阈值的使用,这类方法存在主观性和鲁棒性不足的问题。为解决上述问题,提出一种基于大规模正脸数据驱动的集成学习算法。通过构建大规模正脸类别,解决正脸姿态分类中类内方差大而类间方差小的问题,避免人为确定阈值带来的主观性问题。利用人脸特征中的姿态信息和大规模集成的方式,区分正脸图像和非正脸图像,提高分类能力,增强鲁棒性。实验结果表明,所提方法无需依赖关键点标注,具有较短的推理时间,在公共数据集上实现了正脸姿态估计。在光照变化、配饰遮挡、小角度和大角度的真实数据集上展示了高效的分类能力。Existing frontal pose estimation tasks rely on the estimation of the 3D pose of the head and the use of empirical thres-holds.This type of method has problems of subjectivity and insufficient robustness.To solve the above problems,an ensemble learning algorithm driven by large-scale frontal face data was proposed.The problem of large intra-class variance and small inter-class variance in frontal face posture classification was solved by constructing a large-scale frontal face category,and the subjectivity problem caused by artificially determining thresholds was avoided.The posture information in facial features and large-scale integration was used to distinguish frontal and non-frontal images,classification capabilities were improved,and the robustness was enhanced.Experimental results show that the proposed method does not need to rely on key point annotation,has short inference time,and achieves frontal pose estimation on public data sets.Efficient classification capabilities are demonstrated on real data sets with illumination changes,accessory occlusion,small angles and large angles.

关 键 词：正脸姿态估计 表征信息 最近邻分类 集成学习 余弦距离 机器学习 大规模数据 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]