A new intelligent retail container system with a dual neural network model design

Abstract

In recent years, image recognition technology based on deep learning has become the main solution for intelligent retail containers (IRC). This article introduces a new intelligent retail container system with a dual neural network model. Compared with the previous single-model design, the new one has significantly improved its detection recall and classification accuracy besides reducing greatly the model's retraining time caused by the increasing in new retail varieties. First, using the Faster RCNN model to complete the rough detection of retail categories (classified by outer package) to improve the detection recall; second, using the ResNet50 model to complete the fine classification of retail subcategories (classified by goods variety) to promote classification accuracy. At the same time, a variety of ablation experiments are carried out on the hard samples set of our project by means of several data augments. Some design methods and practical experience proposed in this article can be helpful for the CV (computer vision) incubation projects in the landing stage.