Robust Visual Product Matching in Intelligent Warehouses via Siamese Networks and Adaptive Data Augmentation

Abstract

In order to meet the ever-changing demands and higher accuracy requirements of intelligent warehouse management systems, automatic product identification technology will be added. The lack of stable visual matching in large-scale logistics is the main issue this paper aims to address. Lighting, occlusion, or any changes to the product can reduce the stability of the solution. Propose an enhanced Siamese network structure while establishing an augmented data pipeline to simulate warehouse changes. Strict labeling and partitioning rules have already been adopted, and a large number of real warehouse images have been used for system validation. In conditions of poor lighting or partial occlusion, the average exceeds 91%, with inference speeds below 50 milliseconds on typical hardware. Category-level accuracy and overall robustness have significantly improved compared to traditional convolutional and metric learning baselines. Adaptive decision logic and context-aware network training can reduce failure rates and ensure stable operation of visually similar or label-ambiguous products. The aforementioned findings lay the foundation for a highly scalable platform for the next generation of warehouse automation. To scale, continuous learning and various data types will be added in the future.