基于移动平台的快速相似脸检索

主要研究了移动平台上的相似脸检索问题.对于移动端,首先采用基于稀疏约束的级联回归模型进行精确的人脸配准,该方法不但能够筛选鲁棒的特征,而且可以将模型的大小压缩到原来的5%左右.接着,在某些关键点周围提取高维的纹理特征,并通过稀疏投影降维.对于服务器端,采用级联形状和纹理特征的方式进行高效的相似脸检索.首先基于稀疏形状重构的方式筛选脸型相似的人脸,然后基于稀疏纹理重构的方法确定相似脸.在三星Note 3智能手机上,人脸图像的配准时间约10 ms.在扩展的LFW(Labeled Face in Wild)数据库上,相似脸检索时间约1.5 s,整个模型大小约5.4 MB.大量实验结果表明,配准方法精度高,速度快,模型小;相似脸检索的方法效率高,检索结果更符合人们的视觉感受.

Fast similar face retrieval based on mobile platform

The problem of similar face retrieval on the mobile platform was studied. For the mobile terminal, sparse constrained cascade regression model was utilized to align the face image accurately, which could not only select the robust features, but also compress the model size to about 5% compared to the original size. Then high-dimensional texture features were extracted around some specific landmarks, and compressed by sparse projection. For the server side, shape and texture features were cascaded to retrieve similar faces efficiently. Faces with similar facial shape were selected by sparse shape reconstruction, and similar faces were finally selected by sparse texture reconstruction. On the Samsung smart phone of Note 3, the alignment time for each face image was about 10 ms. On the extended labeled face in wild (LFW) database, the retrieval time was about 1.5 s and the size of the whole model was only 5.4 MB. Extensive experiments show that the proposed alignment method is accurate and fast with compact model size. Similar face retrieval is efficient and the results are consistent with human visual perception.