Active Learning Based on Prototype Vector Feature Adaptive Fusion

From successfully embedding the KNN classifier into a network using prototype vectors, to exploring various active learning methods in deep learning to find uncertain and representative samples for active learning, feature fusion in the learned feature space has been continuously investigated. In this paper, we aim to further combine prototype vectors with feature fusion and propose an active learning based on prototype vector feature adaptive fusion. Firstly, we explore the features of prototype vectors in the NCE-Net model, where they represent common features for each class. Secondly, we construct a prototype vector matrix based on the feature space. Then, we fuse the prototype vector matrix with the feature matrix of all unlabeled samples and add the samples with inconsistent pseudo-labels before and after feature fusion to the candidate set. Finally, we cluster the candidate samples to consider the representativeness of the selected samples in active learning. Experimental results on classic datasets for image classification show that the proposed method achieves higher accuracy compared to traditional active learning methods.