Abstract

In linear representation-based image classification, an  unlabeled sample is represented by the entire training set.  To obtain a stable and discriminative solution,  regularization on the vector of representation coefficients is  necessary. For example, the representation in sparse  representation-based classification (SRC) uses L1 norm  penalty as regularization, which is equal to lasso. However,  lasso overemphasizes the role of sparseness while ignoring  the inherent structure among samples belonging to a same  class. Many recent developed representation classifications  have adopted lasso-type regressions to improve the  performance. In this paper, we propose the adaptive class  preserving representation for classification (ACPRC). Our  method is related to group lasso based classification but  different in two key points: When training samples in a  class are uncorrelated, ACPRC turns into SRC; when  samples in a class are highly correlated, it obtains similar  result as group lasso. The superiority of ACPRC over other  state-of-the-art regularization techniques including lasso,  group lasso, sparse group lasso, etc. are evaluated by  extensive experiments.