A Hybrid Artistic Model Using Deepy-Dream Model and Multiple Convolutional Neural Networks Architectures

Abstract

The significant increase in drug abuse cases prompts developers to investigate techniques that mimic the hallucinations imagined by addicts and abusers, in addition to the increasing demand for the use of decorative images resulting from the use of computer technologies. This research uses Deep Dream and Neural Style Transfer technologies to solve this problem. Despite the significance researches on Deep Dream technology, there are several limitations in existing studies, including image quality and evaluation metrics. We have successfully addressed these issues by improving image quality and diversifying the types of generated images. This enhancement allows for more effective use of Deep Dream in simulating hallucinated images. Moreover, the high-quality generated images can be saved for dataset enlargement, like the augmentation process. Our proposed deepy-dream model combines features from five convolutional neural network architectures: VGG16, VGG19, Inception v3, Inception-ResNet-v2, and Xception. Additionally, we generate Deep Dream images by implementing each architecture as a separate Deep Dream model. We have employed autoencoder Deep Dream model as another method. To evaluate the performance of our models, we utilize normalized cross-correlation and structural similarity indexes as metrics. The values obtained for those two quality measures for our proposed deepy-dream model are 0.1863 and 0.0856, respectively, indicating effective performance. When considering the content image, the metrics yield values of 0.8119 and 0.3097, respectively. As for the style image, the corresponding quality measure values are 0.0007 and 0.0073, respectively.

Keywords

• Deep dream image
• neural style transfer
• gram matrix