Few-shot FMCG Product Recognition based on Feature Transfer Learning

Fast-moving consumer goods (FMCG), such as beverages, ice cream, and tobacco, have to respond to market demands in a timely manner and carry out fast logistics. It is very time-consuming and labor-intensive to guarantee the quantity of goods at every check point, including inbound and outbound, transportation, warehousing, sorting, and delivery, where all items are checked and counted. Thanks to the rapid development of deep learning, detection, and recognition algorithms are widely used in object recognition. However, for some new products on the market, only a few design photos are available for training, a problem known as a few-shot learning case, in which the recognition accuracy will drop significantly. To address this problem, this paper proposes a method based on feature transfer learning (FTL). This method applies a pre-trained model to extract features as the basis representation, which is transferred in feature space through a feature transfer model, trained on samples with design and realistic samples. The new feature representation is applied in real scenarios for product recognition. We build a dataset of FMCG with tobacco, beverage, and ice cream to validate the proposed method. Experiments demonstrated that in the few-shot setting our proposed method is able to boost the recognition accuracy from 66% (without FTL) to 97% (with FTL).