Deep CNN Feature Resampling and Ensemble Based on Cross Validation for Image Classification

Deep convolutional neural networks (CNNs) such as AlexNet, VGGNet, ResNet, EfficientNet, and MobileNet have been extensively employed in image classification tasks. A common solution is directly feeding deep CNN features extracted from a deep network into a classification function. However, this solution may easily result in poor accuracy and robustness due to the single experimental result. One alternative is utilizing an ensemble of multiple deep networks. And this would bring very expensive, even unacceptable computational complexity. Thus, we propose a new deep CNN feature ensemble frame based on multiple cross validation resampling results of the single feature layer to cope with the above two issues. Theoretically, the proposed method is proved that having a smaller error rate than the single feature layer method and the same Rademacher complexity as the single feature layer method. Moreover, extensive experiments on several challenging image classification databases demonstrate the superiority of the proposed method.