科研之友
微信
新浪微博
Facebook
分享链接
摘要
In a general face recognition scenario, classifications attend to assign a label to a single probe image. So far a branch of classification methods, which assume that a probe image tends to lie on the same class-specific subspace as the gallery images from the same class, have drawn wide attention for their good performance. Actually, those linear regression based classifications are sufficient to achieve promising recognition accuracy. However if there are wide ranges of variations on probe images such as pixel noises, lighting variant, they could deviate the probe images from their correct locations in feature space. To solve this problem, we propose a new linear regression based method by generating an extended set for a probe image. In the first step of our method, we not only produce the low dimension features for a probe but also generate virtual samples by adding randomness into downsampling. The second step is to classify the probe by using canonical correlation analysis. As the generated virtual probe samples have high possibility to cover the correct location in feature space, our proposed method shows promising performance in the experiments.
