Image classification using regularized convolutional neural network design with dimensionality reduction modules: RCNN–DRM

Deep Learning is one of the machine learning area, which is widely used in recent research fields. In this, the work exhibits about working of the Convolutional Neural Networks (CNNs) for image classification. Deep learning approaches are better than the traditional learning algorithms when the data size is large because every day, a vast volume of data is accumulated everywhere. In deep learning, Convolutional Neural Network is one of the leading architecture. Convolutional Neural Network contains pre-trained models to transfer knowledge for learning the features, and such models are LeNet, AlexNet, GoogleNet, VGG16, VGG19, Resnet50, etc. These architectures are trained with a large ImageNet dataset, which contains millions of images. Moreover, these trained networks are also used to do new tasks. Among these pre-trained models, GoogleNet has less number of parameters, and this causes to reduce the computation complexity. We propose a deep network with Dimensionality Reduction Module (DRM), which works on less training data, and produce more accurate classification with minimum processing time and also a minimum number of parameters with regularization. The performance of classification, as well as training time and classification time of the proposed architecture, is measured with popular datasets such as ORL, Adience face dataset, Caltech101, and CIFAR10. The proposed architecture achieves better performance with less time when compared with the state of the work.