Graph feature representation network on image compression

• Main Author: Xing, Ziyan
• Other Authors: Anamitra Makur
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Image compression; Graph representation; Deep learning method; End-to-end compression; Convolutional block attention module
• Online Access: https://hdl.handle.net/10356/175527
• Institution: Nanyang Technological University

This dissertation investigates image compression algorithms integrating deep learning techniques. Due to the popularity of internet technology and smart devices, the need for image signals as the primary medium for information transmission is growing at an alarming rate, so a new compression technique is proposed that aims to increase the compression rate and reduce the transmission cost while ensuring the image quality. The main objective of this dissertation is to explore, analyse and improve the existing image compression techniques and to present the design concept, network structure and working principle of GFRnet based compression network. Experimental design, model training and testing using MS COCO dataset, and results and comparative experimental analysis were conducted. Although this project achieved more satisfactory experimental results in terms of experimental design, feature vector visualisation and its analysis of the effect on the output image, there are also problems of colour distortion and unsatisfactory visual effects. The dissertation also explores the challenges and problems encountered in the research and suggests possible improvement options.