Few-shot Learning with Attention Mechanism and Transfer Learning for Import and Export Commodities Classification

Abstract

As deep learning theory develops rapidly, the convolutional neural network model has been widely used in many fields with its powerful characterization ability and outstanding classification performance. Therefore, the number of parameters in deep convolutional neural network models is usually very large, and massive labeled data is often required for model training. In some scenarios, it is difficult or even impossible to collect enough labeled data. Instead, few-shot learning can obtain considerable learning performance with a small sample size. Thus, we study a few-shot learning model with feature enhancement and transfer learning on a small dataset of import and export commodities. We choose ResNetl 8 as the backbone and use data augmentation to expand the original small dataset before training, which somewhat alleviates the overfitting problem of the convolutional neural network model. Moreover, we introduce the attention module and transfer learning into the backbone. The experimental results on the dataset clearly verify the effectiveness of above methods.