查看文章

Product image matching is essential on e-commerce platforms since the target customers have to distinguish the same products organized in varied manners by different companies. Machine learning methods can be applied for such kind of task, which treat it as a classification problem. However, it is intrinsically intractable to enforce the model to expand to new products due to its invisibility to unknown categories. Metric learning has emerged as a promising approach to address this issue, which is designed to measure the distances between data points. In this paper, a metric learning based vision transformer (ML-VIT) is proposed to learn the embeddings of the product image data and match the images with small Euclidean distances to the same products. The proposed ML-VIT adopts Arcface loss to achieve intra-class compactness and inter-class dispersion. Compared with Siamese neural network and …