Detection of Ginkgo biloba seed defects based on feature adaptive learning and nuclear magnetic resonance technology

In order to solve the problem that artificial selection of high-quality seeds is limited to the external characteristics of seeds and inefficient, this paper develops a new method to identify the internal defects Ginkgo biloba seeds by low-field nuclear magnetic resonance (LF-NMR) technology and convolution neural network. First, 4500 images reflecting the internal characteristics of ginkgo seeds are collected by low-field nuclear magnetic resonance system. Region of interest segmentation, mean enhancement algorithm, and proportional enhancement algorithm are used to preprocess the magnetic resonance imaging (MRI) of Ginkgo biloba seeds. Then, the Alex-Net model is improved to create a new SE_AlexNet_MiniConv (SAMC) model, including one depth texture network and three attention modules of the squeeze–excitation network. Furthermore, transfer learning is applied to improve the model performance of Ginkgo biloba seeds classification. Lastly, adjustments are made to the network structure, and the SAMC model hyper-parametric is adjusted. The data set is divided into training set, verification set, and test set according to the ratio of 7:2:1. The result shows that the classification accuracy of SAMC model for three kinds of defects (rotten, normal, and embryo absent) reaches to 96.92%, which is higher than other four common models, including Alex-Net, VggNet-16, ResNet-18, and ResNet-50. In addition, gradient class activation mapping (Grad-CAM) is used to analyze the attention area of the SAMC model for each category. This research demonstrates that MRI based on feature adaptive learning achieves high precision, recall, and F1 scores. LF-NMR technology is feasible and effective for detecting internal defects of Ginkgo biloba seeds non-destructively.