An Experimental Analysis of Predictive Coding Based On Artificial Neural Networks for Image Decoding
Open Access
- Author:
- Sun, Yanbo
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 26, 2018
- Committee Members:
- Clyde Lee Giles, Thesis Advisor/Co-Advisor
Jesse Louis Barlow, Committee Member - Keywords:
- Artificial Neural Network
Image Decoding
Predictive Coding
Spatial Redundancies - Abstract:
- Prediction techniques have been extensively studied for images and video compression over the last few decades. Currently, most existing images and video compression solutions contain some predictive encoding in their algorithms. For example, it is well known that the 25-year history of the joint image group (JPEG) image coding standard predictive encodes quantified DC transform coefficients. There some other sorts of predictive coding techniques which implement redundancies inside images and videos. Similarly, in some particular cases such as image sequences, the compression system usually uses the motion compensation prediction to utilize the time redundancy. We discuss some of the most essential predictive coding techniques used in state-of-the-art image and video encoders. Then we provide some basic concepts about video color spaces representation. Various technologies including linear and non-linear predictions methods are presented in the following part. In this thesis, we have implemented and presented an Artificial Neural Network Approach for the encoding of JPEG images. We have shown that JPEG compression can be significantly improved with our decoding method. In these experiments, we have implemented MLP and RNN models on tiny-imagenet data sets. We performed stateless and stateful prediction without using spatial information as well as predictions using partial spatial information and full spatial information. We use the partial and full spatial redundancy in the image to implement a nonlinear mapping function. Compared with the original image, the proposed method outperforms the JPEG image in generating predictive images with relatively small distortion.