Enhancing performance of conventional image codecs using CNN based image sub-sampling and super resolution

Abstract

The goal of image compression is to reduce the number of bits required to represent an image with a minimum loss of visual quality. However, conventional image compression algorithms, such as JPEG, produce unpleasant artifacts in decoded images at high compression ratios. This thesis investigates a CNN-based approach to improve the performance of the widely used JPEG codec. In recent times, there has been an increasing interest in using convolutional neural network (CNN) for various image processing tasks, owing to their ability to learn very compact features from images. However, the use of CNNs to improve the performance of existing image codecs is very limited in literature. Motivated by this, we investigate a CNN based image compression framework which improves the performance of the JPEG algorithm by optimally sub-sampling the input image with a CNN referred to as compact convolutional neural network (ComCNN) prior to JPEG encoding and by performing super resolution and enhancement of the decoded image with a CNN referred to as enhancement based reconstruction convolutional neural network (EBR-CNN). Both CNNs are optimally trained to minimize the end-to-end image distortion for a given value of the JPEG quality factor. Experimental results are presented which compare the performance of the proposed compression framework with several alternative learning and non-learning based image sub-sampling and super resolution methods. These results show that the proposed method provides noticeable improvements in decoded image quality compared to the other alternatives.