A multi-view images classification based on shallow convolutional neural network

Abstract

Multi-view images represent the target object from multiple perspectives. Learning the target object information from different viewpoints helps to improve the accuracy of multi-view images classification. We propose a multi-view recognition method (SCNN-a) based on shallow convolutional neural network. In order to improve the generalization capability and classification performance of model, we develop a new multi-view images classification method (SCNN), which adds Dropout (after each max-pooling layer) technology to SCNN-a. SCNN-a and SCNN regard the images from predefined views as latent variables, and extract the high-order features of multi-view images with two convolutional layers. Experimental results show that SCNN achieves similar accuracy to the state-of-the-art result of [7] with less layers and time complexity.