University Archives

This study presents a deep learning framework for Large-Scale Product Retrieval, critical for modern e-commerce. Identifying items from massive databases is challenging due to fine-grained differences and viewpoint variations. The methodology implements a high-performance pipeline tailored for the Stanford Online Products (SOP) dataset. In the first stage, Feature Extraction, a ResNet-50 model with 224x224 resolution generates robust visual descriptors. The second stage, Feature Aggregation, utilizes a learnable Generalized-Mean (GeM) pooling layer to capture discriminative details. The architecture further incorporates an end-to-end whitening layer and contrastive loss with dynamic hard-negative mining to improve representation learning. The resulting system takes a query image as input and retrieves visually similar products from the database.