基于小波的训练样本增强的单样本人脸识别

Wavelet-based Training Sample Enhancement for Face Recognition with One Training Sample

目前有许多处理正面人脸的识别方法,当有充分数量的有代表性的训练样本时,能取得较好的识别效果。然而当每个人只有一个训练样本时,这些方法的识别性能则会下降。文章提出了一种基于小波分解低频子带的训练样本增强的方法,为了加强单样本的分类信息,将训练样本与其小波低频子带的重构图组合成为增强样本,然后在训练集的平均频谱图像的奇异值分解的统一特征空间进行识别。在Yale人脸库上的实验结果表明,当训练集中每个人只有一幅人脸图像时,该文提出的方法比统一特征空间奇异值分解方法取得更高的识别率。

At present,many methods can deal well with frontal view face recognition when there is sufficient number of representative training samples.However,the recognition performance of these methods decreases when only one training sample per person is available.In this paper,we propose an enhancement method of training sample based on Wavelet Transform Low-Frequency Band （WTLFB）.In order to enhance the classification information of single training sample, each training sample is combined with its reconstructed image based on WTLFB into an enhanced sample.Then recognition is performed on a uniform eigen-space that obtained from Singular Value Decomposition （SVD） of the mean spectrum image of the enhanced training set.Experimental results show that on the Yale database where each person has only one training sample,the recognition accuracy of the proposed method is higher than the uniform eigen-space SVD method.