Classification of commodity items by their images using convolution neural networks

The paper considers stage of determining the full HS code of the target group "shoes” by considering images of goods, attached to electronic documentation. Reasons of using convolutional neural networks (CNNs) for image classification is presented. Possible approaches to constructing specialized neural network classifiers are examined. A comparative analysis is conducted on the effectiveness of methods based on fine-tuning pre-existing classifiers (transfer learning) versus building convolutional networks trained exclusively on labeled data from a selected product assortment. The challenges of acquiring training datasets through scraping specialized websites and generating dataset elements via artificial intelligence systems specialized in on-demand image generation are explored.