A Parallel Feature Expansion Classification Model with Feature-based Attention Mechanism

Because of the close relationship between artificial neural network and neuroscience, some visual mechanisms are often used to improve the performance of convolutional neural networks (CNNs). Inspired by parallel processing of human brain visual information and information fusion in common brain regions, this paper designs a parallel feature expansion model (PFEM). The model can extract two features based on a parallel CNN structure, and performs two quadratic term transformations for feature expansion at the end of the feature extractors, then all the features are input to the fully connected layers and classifier after fusion. We further add feature-based attention to PFEM to correct the activation values of CNN feature maps. Experimental results on CIFAR-10 dataset show that PFEM with feature-based attention can improve the classification accuracy of the CNN.