| Abstract [eng] |
The research aims to develop a self-learning, barcodeless product recognition system for self-checkout services in food retail. To achieve this, the study focuses on several objectives. First, it involves analyzing self-checkout product images to create a schema for training neural networks, addressing challenges such as empty images, customer interference, products in bags, and sales imbalances. Second, the research proposes methods for evaluating image quality and image emptiness, testing their effectiveness, and developing a neural network architecture for product classification. This architecture is designed to work efficiently on low-powered, GPU-less machines and is compared with existing state-of-the-art systems. Lastly, the study suggests a method for grouping similar products to improve prediction accuracy, with an emphasis on increasing accuracy beyond the traditional top-1 approach. Overall, the research seeks to enhance the efficiency of self-checkout systems, even with limited computational resources. |
