Abstract
With the rapid growth of big data applications, the training process of deep neural networks is getting more expensive in terms of the computational cost. In this paper, we propose an algorithm to exploit the reliability of existing convolutional neural networks that has been gained during earlier training processes. We use fuzzy integrals to perform late fusion on the classification decisions taken by pre-trained classifiers. The proposed method was evaluated using the ImageNet benchmark with ten different pre-trained state-of-the-arts Convolutional Neural Networks (CNN) models. The evaluation results show that the proposed fuzzy-based fusion method could achieve better performance than the best of the contributing models, in terms of recognition accuracy. The accuracy improvement ranges from 8% to 30% better than the used pre-trained classifiers.