Dictionary based Bayesian Classifier Learning

Abstract

Dictionary and Classifier learning with discriminatory and joint behavior is a considerably effective area in ML research being applied particularly for face recognition, action recognition, and object detection. We present an approach to improve classification performance by enhancing joint learning of the dictionary and classifier. Dictionary and classifier are separately or jointly learned with different sparse representations for training and labels' data. At the perdition stage, sparse representation of a test sample computed over the learned dictionary is used as input for the classifier for classification. The accuracy of the classifier can be increased by using sparse representations of labels over the classifier. To mitigate this issue, we present an approach to jointly learn the same representations for both the test samples and the corresponding labels. At the prediction stage, the computed representation of a test sample over the dictionary will serve the purpose. We performed tests to confirm the effectiveness of our approach, using the Gibbs sampler as an inference for face, object, scene, and action recognition. We compared the results also with other state-of-the-art approaches in the area of Dictionary and Classifier learning. Our approach achieves a classification accuracy significantly higher than that of other approaches.