The Effect of Changing Targeted Layers of the Deep Dream Technique Using VGG-16 Model

Abstract

The deep dream is one of the most recent techniques in deep learning. It is used in many applications, such as decorating and modifying images with motifs and simulating the patients' hallucinations. This study presents a deep dream model that generates deep dream images using a convolutional neural network (CNN). Firstly, we survey the layers of each block in the network, then choose the required layers, and extract their features to maximize it. This process repeats several iterations as needed, computes the total loss, and extracts the final deep dream images. We apply this operation on different layers two times; the former is on the low-level layers, and the latter is on the high-level layers. The results of applying this operation are different, where the resulting image from applying deep dream on the high-level layers are clearer than those resulting from low-level layers. Also, the loss of the images of low-level layers ranges between 31.1435 and 31.1435, while the loss of the images of upper layers ranges between 20.0704 and 32.1625.