Few-shot Learning for Quality Detection of Logistical Parcels

With advancements in science and technology and the ongoing enhancements made to distribution systems, the volume of logistical parcels being handled has increased significantly. However, the quality control process for logistical parcels still relies on manual inspections, which proves to be inefficient, especially when attempting to flag damaged parcels, an issue that requires immediate attention. Moreover, as the percentage of damaged parcels within the overall volume is relatively low, this presents a challenge that can be addressed through few-shot learning techniques. In this study, we present an application of few-shot learning to the recognition of logistical parcels. We compare three distinct few-shot learning methods — Model-Agnostic MetaLearning (MAML), Siamese Neural Network, and DAFT, and identify the optimal approach through the evaluation of our self-designed parcel dataset. Results from our experiments reveal that the MAML method is not effective for parcel recognition. The Siamese Neural Network method attained a high accuracy of 70.31% on small-scale parcel datasets while the DAFT method achieved the best accuracy of 83.33% in identifying damaged parcels. Consequently, the DAFT approach provides an excellent recognition solution for logistical parcels and empowers automated damage parcel identification with improved performance.