CSClust: A Luxury Goods Image Authentication Algorithm Based on Contrastive Clustering

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.