A Metric Learning Based System for Retail Product Recognition and Novel Class Discovery

Developing a retail product recognition system presents different challenges compared to traditional object recognition systems. Retail products contain classes that are very similar to each other but differ in small details. The current product classes in the system must be frequently expanded with new product classes or updated with packaging changes. For these reasons, we developed a metric learning-based product recognition system that provides superior product recognition performance and the ability to add new products and update existing ones. The key components of our system are the ConvNext_nano feature extractor, ArcFace loss, feature vector database, similarity search, and clustering algorithm. Through tests performed on a custom dataset consisting of 775,000 samples from 994 classes, we achieved an accuracy of 93.7% in product recognition and an average clustering accuracy of 67.58% in discovering new products.