Interactive and discriminative analysis dictionary learning for image classification

Dictionary learning is widely utilized in pattern recognition, and analysis dictionary learning is a prevalent image classification method. However, its classification performance still has much room for improvement. Constructing the powerful discriminative constraint by exploiting the intrinsic characteristics of sample is an effective approach for enhancing the performance, thus how to design an effective constraint is a problem worth studying. On the other hand, some analysis dictionary learning models incorporate the classification error constraint. However, this constraint always adopts a strict binary matrix as the target matrix, which is harmful to the improvement of classification performance. To solve these issues, we propose an interactive and discriminative analysis dictionary learning for image classification, The ordinal locality preserving technique is utilized to to preserve the topology information of the samples, and Fisher constraint is applied to promote the discrepancy of inter-class coding coefficients and the similarity of intra-class coding coefficients. Furthermore, the target matrix is relaxed by exploiting \(\ell _{21}\) ${\ell }_{21}$ norm, which can provide more freedom to the classifier parameter. Finally, comparative experiments on different types of data sets show the efficacy of the proposed method, For 15 Scene dataset, the proposed IDADL improve the classification accuracies by 1.5%, 0.9% and 0.6% over FDDL, SLCADL and RADPL, respectively. Besides, the performance of the proposed model presents superiority over some classical deep models, For CMU PIE, the proposed IDADL promote the classification accuracies by 4.3% and 1.6% over AlexNet and VGG, respectively.