Deep learning based automated non-barcoded product identification system for in-person shopping

Abstract

Deep learning, an advanced extension of machine learning, is widely used for complex challenges. This study presents an automated system for recognizing non-barcoded products in retail environments, addressing inefficiencies in manual identification. A novel dataset, MAKBUL, was developed with 4500 RGB images from 30 product categories (150 images each), captured using a custom-designed camera system on a retail weighing platform. Experiments with six pre-trained CNNs showed that DenseNet201 achieved the highest individual accuracy (96.00%) when fine-tuned with a 0.00001 learning rate, RMSProp optimizer, and 100 epochs. Using feature extraction, EfficientNet + SVM reached 96.07% accuracy. The best result, 96.56%, was obtained through feature-level fusion, combining features from all six models. The study tested 27 hyperparameter combinations and applied 5-fold cross-validation for robustness. Findings highlight the efficiency of transfer learning in multi-class product recognition, reducing customer wait times and manual labor. This research lays the groundwork for future automated retail systems, with potential enhancements including larger datasets, varied lighting conditions, and additional product categories.