E-commerce image recognition using attention model and YOLACT neural network

Abstract

Objectives. We propose the algorithm for e-commerce image recognition using attention model and neural network YOLACT. A modular architecture is used that applies an attention model to different branches of the network in order to improve the interaction between image cross-features.
Methods. The main methods to recognize e-commerce products are the creation and annotation of a dataset for the neural network training, the choice of architecture and embedding an attention model, the validation and testing, and interpretation of the results.
Results. Convolutional neural network YOLACT has been modified by the attention model to solve image recognition task that allowed to obtain results superior in quality to the results showed by classic YOLACT.
Conclusion. In the course of the experiment, a data set of e-commerce products was prepared, annotated, and two neural networks were built to compare the results. The results of the study showed that the use of the attention model has a positive effect on both the quality of the trained network and on the rate of convergence, which is reflected in improved metrics for object recognition and segmentation.