Feature transforms for image data augmentation

Abstract A problem with convolutional neural networks (CNNs) is that they require large datasets to obtain adequate robustness; on small datasets, are prone overfitting. Many methods have been proposed overcome this shortcoming CNNs. In cases where additional samples cannot easily be collected, a common approach generate more data points from existing using an augmentation technique. image classification, many approaches utilize simple manipulation algorithms. work, we propose some new for based several transformations: the Fourier transform (FT), Radon (RT), and discrete cosine (DCT). These other considered in order quantify their effectiveness creating ensembles of networks. The novelty research consider different strategies training sets which train classifiers combined into ensemble. Specifically, idea create ensemble kind bagging set, each model trained set obtained by augmenting original approaches. We build level adding images generated combining fourteen approaches, three FT, RT, DCT, here first time. Pretrained ResNet50 finetuned include derived method. fusions evaluated compared across eleven benchmarks. Results show building produce not only compete competitively against state-of-the-art but often surpass best reported literature.