Open-set pattern recognition in computer vision

• Main Author: Liang, Yuqian
• Other Authors: Mao Kezhi
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; CNN; OpenMax; Open-set
• Online Access: https://hdl.handle.net/10356/178226
• Institution: Nanyang Technological University

In traditional classification or recognition tasks in the field of Computer Vision, people assume that all test sample classes are first encountered during training, which is unrealistic. Open Set Recognition aims to reject unknown classes and correctly classify known classes during testing. This project aims in studying and implementing the application of Open Set Recognition technology in image classification tasks. This paper first reviews the proposal and development of Open Set Recognition technology, and then focuses on the study of five classic Open Set Recognition algorithms. Based on self-defined convolutional neural networks, this article attempts to replicate the algorithms of four of these methods. We conducted experiments on these four algorithms based on mainstream open dataset settings. The results indicate that compared to traditional deep neural network based classification algorithms as the baseline, the four algorithms studied in this dissertation have successfully reduced open space risks and achieved obvious advantages in performance in open set image recognition tasks.