A Multiscale Interactive Attention Network for Recognizing Camellia Seed Oil with Fuzzy Features

Abstract

The adulteration of camellia seed oil with different processes will seriously violate the rights and interests of consumers. The accurate identification of camellia seed oil processes is of great significance to reduce such illegal activities. However, the fatty acid composition of camellia seed oil is complex and the content varies greatly in the same process, while the difference is small in different processes. This multivariate data are easy to lead to the fuzzy characteristics of camellia seed oil, which increases the difficulty of identifying camellia seed oil quality. To solve these problems, we propose a multi-scale interactive attention network (MIANet) for the accurate identification of camellia seed oil. Firstly, a one-dimensional multi-scale convolutional feature extraction method (OMCM) was proposed, which was used to reduce the difference from multivariate fuzzy features and better solve the problem of fuzzy features of camellia seed oil fatty acids with the same process. Secondly, the interactive attention mechanism (IA) was proposed to enhance the deep characteristics of multivariate fatty acids from the fusion of two dimensions, so that the model paid more attention to the subtle differences between different processes, and effectively solved the problem of fuzzy fatty acid characteristics of camellia seed oil in different processes. Finally, in order to verify the effectiveness of MIANet, MIANet is compared with classical machine learning methods such as SVM, KNN, LR, LDA, QDA, classical deep learning method AlexNet, and the most advanced deep learning methods such as DMCNN and HCA-MFFNet. The accuracy of MIANet reached 94.10%, which was better than the eight methods. The experimental results show that MIANet is an effective method for the accurate identification of camellia seed oil data with fuzzy characteristics.