单样本的低分辨率单目标人脸识别算法

针对只给定单幅目标图像的情况下,而要在监控视频中查找出该目标人脸图像的问题,提出了一种单样本的低分辨率单目标人脸识别算法。考虑到待识别样本集中的目标与非目标对象数量严重不均衡,以及单目标问题无法利用不同类别间的互斥关系。首先在待识别样本集中,通过聚类算法,将单目标的识别问题转化为多目标识别问题,进而提高开集人脸识别算法的鲁棒性;其次,利用迭代标签传播算法不断优化待识别样本的归属类别;在迭代过程中,按照置信概率估计每个类别的人脸确认阈值,以解决单样本无法训练分类器的问题。在多个人脸数据集上的实验结果表明,该算法对于单目标的单样本的人脸识别精确率既能逼近100%,也具有较高的召回率。

Low resolution single object face recognition algorithm with single sample

Recognizing the target face from the surveillance video using the given single target image is still a challenging issue in practical applications. Therefore, a low resolution single target face recognition algorithm with single sample is proposed in this work. Two significant limitations are taken into consideration, the sample numbers of the target objects and non target objects are seriously unbalanced in the probe set, and the single object problem cannot utilize the mutually exclusive relationship between different categories. In this paper, firstly, to increase the robustness of the open face face recognition, the clustering algorithm is utilized to transform the single object recognition problem into a multi classification recognition problem. Furthermore, the iterative label propagation algorithm is applied to optimize the attribution probability of the probe sample continuously. During the iteration, the face recognition threshold of each object is estimated according to the confidence probability. Hence, the single sample is capable to train the classifier. Finally, experimental results on multiple face datasets show that the proposed algorithm can achieve good performances in both accuracy and recall rate.