Self-weighted discriminative metric learning based on deep features for scene recognition

Abstract

Deep features have achieved impressive performance for image recognition. However, directly using the pre-trained deep models on scene images is not proper, because of intra-class diversity and inter-class similarity of the scene images. In this paper, based on deep features, a self-weighted discriminative metric learning (SDML) method is proposed to deal with the scene recognition. Specifically, biobjective discriminative metric function is first defined to fully exploit discriminative information from deep features. Second, a self-weighted coefficient is integrated into the biobjective function with the aim of adaptively balancing the discriminative distance between the same class and different classes. Moreover, a nuclear norm regularization is introduced to promote low rank metric. An alternating iterative optimization strategy is devised to solve the proposed method effectively. Finally, the SDML method can be extended to its kernelized version. Experimental results on three scene datasets demonstrate the superiority of our methods over several existing metric learning methods.