DeepDream Algorithm for Data Augmentation in a Neural Network Ensemble Applied to Multiclass Image Classification

Abstract

This paper presents the application of Deep-Dream algorithm for data augmentation applied to images. This algorithm analyzes the image by a trained neural network and hides main features of the image. Importance and place of the features is estimated based on neural network layers output values. The new neural network trained on the processed data are forced to search and train a new set of features in the data, since the known features have been hidden. Trained on original and processed data neural networks are used in an ensemble. Experiments were conducted with a balanced images dataset with 5000 images in 10 classes. Experiment was conducted with a neural network based on InceptionV3 architecture in two variations: with non pretrained weights and with pretrained weights. The neural network received a 256 × 256 pixel image as input. Training was conducted using categorical cross entropy loss function, accuracy metric, Adam optimizer with a learning rate of 0.0001. The improvements of the algorithm are almost insignificant when classifying an ensemble of neural network models for a small number of classes, but the impact of the algorithm increases as the number of classes increases. For binary classification there may be no improvement in ensemble accuracy, but when the number of classes increases and becomes more than 5, the influence of the algorithm on the accuracy of the final ensemble increases. The improvement in ensemble accuracy can be 0.5–4%, depending on the initial training conditions without the use of other types of data processing and augmentation.