CSClust: A Luxury Goods Image Authentication Algorithm Based on Contrastive Clustering
Year: 2024
Type: conference-paper
Abstract: The aim of our paper is to introduce CSClust, a contrastive spectral clustering model for luxury goods image authentication. The approach involves fine-tuning a Vision Transformer (ViT) pre-trained model to enhance representation learning. Utilizing a spectral clustering network, the model embeds latent representations into the feature space and performs subsequent clustering. The training process is conducted in an unsupervised manner on self-generated augmented data, incorporating both instance-level and cluster-level contrastive losses. Experiments have demonstrated that the proposed method exhibits superior accuracy and generalization capability in luxury goods image authentication, and has also achieved state-of-the-art results on the CIFAR-100, ImageNet-10, and ImageNet-dogs datasets.
Authors: Yaxin Chao, Yu Yang, Linna Zhou, Jixin Zou, Min Lei
Institution: Beijing University of Posts and Telecommunications
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4407774683