基于“SOM脸”的选择性单训练样本人脸识别

Finding Important Sub-Areas for Face Recognition from Single Training Image Per Person

基于“SOM脸”计算模型提出一种新的人脸局部区域重要程度度量方法 ,并用于进行选择性单训练样本人脸识别。从机器人脸识别的角度 ,并未预先人为设定任何重要区域 ,而是通过学习来自动发现这些对机器而言相对重要的人脸区域 ,即包含类信息相对丰富的区域 ,并将其进行可视化。实验结果表明 ,在利用了人脸局部区域重要程度信息后 ,识别算法的性能和效率均得到提高 ;特别是仅选择人脸图像中若干部分重要的区域用于识别时 ,在提高识别效率的同时 ,识别性能未见明显下降。

Most subspace methods for the face recognition, such as linear discriminant analysis, discriminant eigenfeatures and fisherface cannot work in the situation of only one training image available per person, while the eigenface technique suffers a great performance drop. A novel method is used to automatically select some important local areas of a face image for face recogn ition with single training image per person. The SOM output space is studied fro m a novel view, where some enlightening ideas from the automatic text analysis f ield are introduced into the face recognition to evaluate the weights of differe nt local areas of face images. The extracted weight information is then used to select some important local facial features for the recognition. Experiments on both the FERET and ORL face database show that the proposed method is more effic ient than its non-weighted counterpart while keeping the recognition accuracy a cceptable. Moreover, the experiment points out that the most important local are a to recognition maybe not those common supposed local areas, such as ear, nose or mouse.