USING DEEP LEARNING FOR IMAGE AND VIDEO COMPRESSION

Abstract

The search for compact representation of images and videos has long been a subject of interest for researchers. A huge amount of multimedia data is shared on the Internet every minute and it is bound to consume large amount of resources. A number of compression schemes already exist like Joint Photographic Experts Group (JPEG), JPEG 2000 i.e. wavelet-based image compression etc but the search for more e cient compression algorithms continues. Deep learning provides us an op- portunity to use it for compression purposes. Recent developments in deep learning have allowed colorization of gray scale images with high accuracy. A recent deep learning based scheme named IdeepColor utilizes Graphics Processing Units (GPUs) to colorize images within seconds in a Linux-based environment. In this research, we study the feasibility of using such deep learning based colorization of images for image compression. The idea is to ignore the color information during encoding and use IdeepColor during decoding. In order to achieve this, three di erent scenarios are proposed and their impact on image compression is studied using di erent image quality assessment methods. In video compression, block matching motion estimation is the most computationally expensive and time consuming process. A recent study has presented a method to predict motion from a single image by using Convolutional Neural Networks (CNN). Using only a single frame, motion of each pixel can be predicted in terms of optical fow. We analyze whether such a method can be used for accelerating the search process for motion vector calculation. Our study reveals that deep learning has the potential to be used for compression purposes.