Image Quality Improvement Using Deep Dream

The term “Deep Dream” was coined by Google engineer Alexander Mordvintsev and originally describes a modified back-propagation which modifies the input of a deep neural network to generate hallucination-like images. This method allows insights to the inner workings of neural networks.

In general, this approach fits the input of the neural network to minimize its objective function—different to standard back-propagation which aims to change the weights of the network to fit the input data. Using this technique the input image could be modified to suit a metric which judges the quality of an image, therefore improving the image.

This topic explores a fun and interesting branch of deep learning which is not researched by the majority of scientists, but still bears interesting albeit unorthodox applications apart from creating hallucination-images. Furthermore Deep Dream gives interesting insights into the inner working of back-propagation and deep learning.

Example image created with Deep Dream techniques

Possible Tasks:

    • Setting up a framework to use Deep Dream for image improvement
    • Training networks for quality assessment
    • Testing different network structures

Examining different image distortions with the framework