Abstract

Background: Retail product recognition is difficult in practice because products often look similar, class distributions are unbalanced, and annotations usually contain only one observed product label, even though products belong to broader category structures. This study addresses this setting as classification under incomplete supervision rather than as a fully annotated multi-label problem. Methods: We evaluate a lightweight hierarchy-aware recognition framework on a laboratory-based retail dataset containing real supermarket products. Visual features are extracted with a pretrained ResNet-50, reduced by truncated singular value decomposition (TSVD), and classified using class-conditional kernel density estimation (KDE). A coarse-to-fine refinement step first assigns each sample to one of seven product groups and then applies a group-specific classifier. Results: The hierarchical TSVD-KDE model achieved 83.7% Top-1 accuracy, 95.8% Top-5 accuracy, and 71.8% macro-averaged F1-score, improving over the flat TSVD-KDE variant and the convolutional neural network (CNN) linear baseline. For rare classes, the model reached 62.4% Top-1 accuracy while maintaining an average inference time of 6.1 ms per sample. Conclusions: The results suggest that combining compact visual representations with probabilistic classification and a simple product hierarchy can improve balanced recognition performance without relying on a large end-to-end architecture.

IPC Classification

Keywords