Department of Image Processing

Following last week's talk about traditional image and video compression codecs, this week we will look into the still relatively young research field, which is using deep learning in the so far very conservative field of image and video compression. Traditionally, all image and video codecs were designed manually in a relatively organized manner. In the last 5 years, however, there has been very active and competitive research into how deep learning can be used to replace the hand-designed transforms in image compression and many partially of fully learned compression methods have been proposed. Current state-of-the-art learned methods can already in many ways compete with or outperform traditional codecs while having several other advantages, such as for example fast and usually cheap domain adaptation or better properties with respect to human perception. In this talk we will review the general principles and requirements of an image compression method, introduce the base structure of learned image codec, and give brief overview into the current state-of-the-art as well as discuss some research challenges. Time permitting, we will look into other related topics, such as targeting realism in image compression or briefly touch learned video compression. Discussion about potential benefits and threats of learned image compression is welcome.