Common visual pattern discovery and analysis

• Main Author: Wang, Zhenzhen
• Other Authors: Tan Yap Peng
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/146898
• Institution: Nanyang Technological University

Given a set of images, common pattern discovery aims to explore the re-occurring patterns based on their similarities and differences. It is a long-standing but challenging task. On the one hand, patterns are frequently occurring visual primitives, and they are present in various forms, such as local features, semantic visual parts or visual objects. On the other hand, there exhibit large variations in visual appearances and structures even within the same kind of visual patterns which are exacerbated by prevalence of mobile-captured images. However, to distinguish visual patterns from one another is fundamental to many tasks in computer vision, such as pattern recognition/classification, object detection/localization, content-based image search. This thesis centers on task-driven common pattern discovery problems and designs several methods based on the characteristics of each task. In the past decades, many studies have attempted to address the problem of visual pattern discovery. Most of them depend on hand-crafted feature representations and step-by-step hierarchically designed strategies, which are difficult to replicate and lack of generalization capability. Thanks to the development of deep learning, many computer vision tasks have evolved to a recorded high accuracy, and the convolutional neural network (CNN) has been involved either as a tool to generate discriminative features or as an end-to-end (e.g., image-to-label) learning model. However, the overwhelming performances are heavily dependent on large-scale high-quality labeled data that are costly to collect. In addition, over-parameterized CNN models which could consume huge computational resources are also critical for impressive performances. Thus, in this thesis we make efforts to explore efficient ways to combine CNNs and common pattern discovery to address the problems in both tasks.