Deep convolutional neural networks based on image data augmentation for visual object recognition

Abstract

Deep Neural Networks (DNNs) have achieved a great success in machine learning. Among a lot of DNN structures, Deep Convolutional Neural Networks (DCNNs) are currently the main tool in the state-of-the-art variety of classification tasks like visual object recognition and handwriting and speech recognition. Despite wide perspectives, DCNNs have still some challenges to deal with. In previous work, we demonstrated the effectiveness of using some regularization techniques such as the dropout to enhance the performance of DCNNs. However, DCNNs need enough training data or even a class balance within datasets to conduct better results. To resolve this problem, some researchers have evoked different data augmentation approaches. This paper presents an extension of a later study. In this work, we conducted and compared the results of many experiments on CIFAR-10, STL-10 and SVHN using variant techniques of data augmentation combined with regularization techniques. The analysis results show that with the right use of data augmentation approaches, it is possible to achieve good results and outperform the state-of-the-art in this field.