Design of a Computationally Economical Image Classifier using Generic Features

摘要

In this paper, we propose an image classification technique which uses a simple autoencoder with a regularizer. Nowadays, Convolutional Neural Networks (CNN) are primarily used for image classification. Our method can be used for image classification with much reduced requirement of computational capability than a complex CNN which has a huge number of degrees of freedom. Here, the terms simple and complex, respectively, correspond to the simplicity and the complexity of a network in terms of the number of learnable parameters (degrees of freedom) and the number of hidden layers. This technique uses features extracted from a pretrained CNN, trained on a completely different dataset. Genetic algorithm solves for the optimal hyperparameters of the pretrained CNN. It is observed that these features serve as important and robust parameters for the training of the autoencoder, as a final average image classification accuracy improvement of about 17.45% is observed with the inclusion of these features. We use a pretrained CNN on MNIST dataset and classify images of several other benchmark datasets. We utilize different classifiers for image classification based on features extracted from the autoencoder and repeat each of the experiments a number of times with different random initialization of the classifier and the weight matrix of the autoencoder. We also perform experiments by pretraining the CNN with different datasets. Our results show a notable image classification accuracy and a significant reduction of training time with respect to a complex CNN.