ConvNet and Dempster-Shafer Theory for Object Recognition

Abstract

We propose a novel classifier based on convolutional neural network (ConvNet) and Dempster-Shafer theory for object recognition allowing for ambiguous pattern rejection, called the ConvNet-BF classifier. In this classifier, a ConvNet with nonlinear convolutional layers and a global pooling layer extracts high-dimensional features from input data. The features are then imported into a belief function classifier, in which they are converted into mass functions and aggregated by Dempster’s rule. Evidence-theoretic rules are finally used for pattern classification and rejection based on the aggregated mass functions. We propose an end-to-end learning strategy for adjusting the parameters in the ConvNet and the belief function classifier simultaneously and determining the rejection loss for evidence-theoretic rules. Experiments with the CIFAR-10, CIFAR-100, and MNIST datasets show that hybridizing belief function classifiers with ConvNets makes it possible to reduce error rates by rejecting patterns that would otherwise be misclassified.