Resumen
Rice is one of the important staple foods for human beings. Germ integrity is an important indicator of rice processing accuracy. Traditional detection methods are time-consuming and highly subjective. In this paper, an EfficientNet?B3?DAN model is proposed to identify the germ integrity. Firstly, ten types of rice with different germ integrity are collected as the training set. Secondly, based on EfficientNet?B3, a dual attention network (DAN) is introduced to sum the outputs of two channels to change the representation of features and further focus on the extraction of features. Finally, the network is trained using transfer learning and tested on a test set. Comparing with AlexNet, VGG16, GoogleNet, ResNet50, MobileNet, and EfficientNet?B3, the experimental illustrate that the detection overall accuracy of EfficientNet?B3?DAN is 94.17%. It is higher than other models. This study can be used for the classification of rice germ integrity to provide guidance for rice and grain processing industries.