Graph-based deep models for image semantic segmentation

• Main Author: Peh, Wei Hang
• Other Authors: Ke Yiping, Kelly
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175317
• Institution: Nanyang Technological University

Image segmentation is a fundamental task in computer vision, classifying each pixel in a image into categories is important to many real world applications. However, acquiring ground truth labels to train deep visual models have been labour-intensive and costly. Recent works have emerged to reduce the cost in training segmentation models using semi-supervised, weakly-supervised, and unsupervised methods. In this study, we use deep features extracted using Vision Transformers and cluster these features using Graph Attention Networks with optimisation objectives from classical graph theory. We apply this proposed method for unsupervised object segmentation and our model yields good performance on the DUTS and ESCCD datasets.