Professor Nang Jong-ho's Research Team Has a Paper Accepted at 'WACV 2025', a Premier Conference in Computer Vision

A paper by the research team of Professor Nang Jong-ho of the Department of Computer Science & Engineering (Kim Jun-tae, master's graduate of the Department of Computer Science & Engineering, and Woo Sung-won, master's student of the Department of Artificial Intelligence (DHE)) has been accepted at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2025), a leading international conference in computer vision. WACV, organized by IEEE/CVF, was held from 28 February to 4 March 2025 at the JW Marriott Starr Pass in Tucson, Arizona, USA.

The paper, titled 'Relational Self-supervised Distillation with Compact Descriptors for Image Copy Detection', proposes RDCD, an efficient method for image copy detection that uses a lightweight network and small feature vectors. Existing image copy detection techniques achieve high accuracy but have been limited in practicality by their large networks and feature vectors.

The study introduces relational self-supervised distillation to transfer knowledge from a large network to a small one. It also applies hard-negative-based contrastive learning to prevent dimensional collapse of the representation space, enabling flexible feature representation even in a small representation space.

The proposed method was evaluated on DISC2021, an international benchmark for image copy detection, recording micro-average accuracy improvements of 5.0%, 4.9% and 5.9% over existing methods at feature vector sizes of 64, 128 and 256 respectively. The research is expected to be usefully applied in building fast, highly efficient image retrieval systems over large image databases.