Confusion-Aware Convolutional Neural Network for Image Classification.

In image classification, it is often encountered that the decision boundaries of some image categories are ambiguous and easy to confuse with each other, thus yielding inferior accuracy on image classification. In this paper, a novel confusion-aware convolutional neural network (CNN) is proposed to address this issue. Different from the coarse-to-fine strategy that has been practiced in existing hierarchical classifiers, our proposed method performs predict-then-correct strategy. At the training stage, a conventional classifier (referred to as the prediction classifier) is trained, and its confusion matrix is estimated by exploiting a cross validation process conducted on the training set. Based on this estimated confusion matrix, a confusion-aware model is then established, and it is used as a decision maker to train a set of correction classifiers for those confusing categories. At the classifying stage, the prediction and correction classifiers collaboratively work together via a hierarchical structure, and the confusion-aware model is used again as a decision maker to select a proper prediction classifier for each confusing category. Experimental results conducted on the Mnist and CIFAR-10 datasets show that the proposed confusion-aware network outperforms the existing CNN classifiers on image classification.