Generating Various Deep Dream Images Through Maximizing the Loss Function of Particular Layers Using Inception-v3 and Inception-ResNet-V2 Models

• Published in: Iraqi journal of science 2024-06, p.3468-3483
• Main Authors: Al-Khazraji, Lafta R., Abbas, Ayad R., Jamil, Abeer S.
• Format: Article
• Language: English
• ISSN: 0067-2904; 2312-1637
• DOI: 10.24996/ijs.2024.65.6.39

     Recently, Deep Learning (DL) has been used in a new technology known as the Deep Dream (DD) to produce images that resemble dreams. It is utilized to mimic hallucinations that drug users or people with schizophrenia experience. Additionally, DD is sometimes incorporated into the images as decoration. This study produces DD images using two deep-CNN model architectures (Inception-ResNet-V2 and Inception-v3). It starts by choosing particular layers in each model (from both lower and upper layers) to maximize their activation function, then detect several iterations. In each iteration, the gradient is computed and then used to compute loss and present the resulting images. Finally, the total loss is presented, and the final deep dream image is visualized. The output of the two models is different, and even for the same model there are some variations, the lower layers' loss values in the Inception-v3 model are significantly higher in comparison to the upper levels' values. In the case of Inception-ResNet-V2, the loss values are convergent.