Jujube Classification Based on a Convolution Neural Network with Multi-channel Weighting and Information Aggregation

Abstract

This paper addresses the quality classification of various types of jujube. That is, for a given type of jujube, we consider how to achieve a highly precise classification of the various dimensions of high-quality jujube, shorten the classification time, improve the efficiency and make it feasible for practical application. The current methods rarely reach the above requirements. To this end, we propose the MI-net (multi-channel weighting and information aggregation) model, which enables the convolutional neural network to learn the aggregate information of the multi-channel feature maps. The proposed model is able to obtain different levels of features and better utilize the channel characteristics, which enhances the discriminative power and generalization ability. The experimental results show that the classification of the full jujube, dried jujube, cracked jujube and broken jujube achieves accuracies of 99.64%, 99.78%, 99.93% and 98.97%, respectively. The overall classification and recognition accuracy rate reaches 99.62%.