A Comprehensive Solution for Deep-Learning Based Cargo Inspection to Discriminate Goods in Containers

In this work, we attempt to classify commodities in containers with HS(harmonized system) codes, which is a challenging task due to the large number of categories in HS codes and its hierarchical structure based on a product's composition and economic activity. To tackle this problem, in this paper we propose an ensemble model which incorporates fine-grained image categorization, data analysis on cargo manifests, and human-in-the-loop paradigm. By employing deep learning, we train a triplet network for fine-grained image categorization. Then, by investigating massive information from cargo manifests, unreasonable predictions can be filtered out. With human-in-the-loop embedded, human intelligence is integrated to justify the resulted HS codes. Moreover, a HS code semantic tree is built to trade off specificity and accuracy.