Naïve sparse representation has stability problem due to its unsupervised nature, which is not preferred for classification tasks. For this problem, this paper presents a novel representation learning method named classification-oriented local representation (CoLR) for image recognition. The core idea of CoLR is to find the most relevant training classes and samples with test sample by taking the merits of class-wise sparseness weighting, sample locality, and label prior. The proposed representation strategy can not only promote a classification-oriented representation, but also boost a locality adaptive representation within the selected training classes. The CoLR model is efficiently solved by Augmented Lagrange Multiplier (ALM) scheme based on a variable splitting strategy. Then, the performance of the proposed model is evaluated on benchmark face datasets and deep object features. Specifically, the deep features of the object dataset are obtained by a well-trained convolutional neural network (CNN) with five convolutional layers and three fully connected layers on the challenging ImageNet. Extensive experiments verify the superiority of CoLR in comparison with some state-of-the-art models.
Local Representation Learning with A Convolutional Autoencoder