基于低场核磁成像的银杏胚检测及分类

银杏胚的有无及发育状况直接影响播种后的发芽率，而这些无法人工检测。该研究基于低场核磁共振成像（Low-field Magnetic Resonance Imaging，LF-MRI）技术，结合深度学习，探索研究银杏种子内部缺陷检测方法。设计了一种基于改进的VGG-16迁移学习图像分类方法，在利用VGG-16卷积层对数据集进行迁移学习时，对模型的学习率、模型结构和归一化3个方面进行改进，比较研究调整初始学习率与更新周期、将全连接层替换为全局卷积层或全局平均池化层并扩展8种微调模型、添加3种归一化层对模型性能的影响。结果表明，全局平均池化层调用方便，能够替代全连接层融合深度特征，提升模型性能；分类输出层置于全局平均池化层后的分类效果较好；组归一化能提升模型性能，并使验证曲线收敛更稳定；学习率对迁移学习模型性能影响极大，使用分段常数衰减策略针对具体应用可有效提高模型性能。相比较VGG-16，该研究构建的基于VGG-Net的银杏种子MR图像识别模型（Ginkgo seed low-field MR images recognition model based on VGG-Net, GV）大小降低89%，参数量降低89%，训练时间减少64%，训练集与验证集损失分别降低64%与45%、准确率分别提高2.4与2.5个百分点，对正常、孔洞、有胚和无胚4个类别的银杏种子平均分类准确率达到97.40%。研究结果为无损监测银杏种子萌发过程、合理预测种子播种后的发芽率提供了思路。

LF-MRI-based detection and classification of ginkgo embryos

Abstract: The presence or absence of an embryo can pose a strong relation with the germination-rate of ginkgo seeds, together with the development status. Unfortunately, neither has been discerned via manual observation without dissection so far. In this study, a non-invasive classification of ginkgo seeds under these internal conditions was explored using deep learning on the low-field magnetic resonance (LF-MR) images. A dataset of four classes was collected, including embryo-present, embryo-absent, normal, and aperture seeds. Each of 1 200 images was sized 32×32 pixels using the LF-MR imaging of 6 000 ginkgo seeds and then categorized according to the dissection evidence. An improved Very Deep Convolutional (VGG-16) network was designed for the ginkgo seed classification (Ginkgo seed LF-MR images recognition model adapted from VGG-Net, or global view (GV)) to classify the LF-MR images, according to whether the presence or absence of embryo in the sagittal plane, or whether the seeds being normal or decayed judging from the coronal plane. The GV reused the convolutional layers of VGG-16 to replace the fully connected (FC) layers with a structural redesign, including a global average pooling layer followed by an FC layer. Compared with the VGG-16, the GV was reduced by 89%, 89%, 64%, 64%, and 45%, respectively, in the size and number of parameters, training time, training loss, and validation loss, indicating an improved accuracy in both training and validation by 2.4 and 2.5 percentage points, respectively. The classification accuracy of ginkgo seed LF-MR images reached 97.40%, and the precision, recall, and F1-score were all above 95%. The reliable detection of internal defects offered a non-invasive approach to identify those ginkgo seeds that cannot germinate. Positive experiences of redesigning and tuning deep learning networks were also gained to classify the LF-MR images, according to the internal defects of ginkgo seeds during the development of GV. At first, the super parameters of both learning rate and update period were determined to perform a transfer-learning of VGG-16 for the LF-MR images classification of ginkgo seeds. Then, a pool of eight candidate structural adaptations was built to test, branching out from the convolutional layers of VGG-16, where two candidates used the FC layers with 3 layers of 512, 512, and 4 neurons, or 2 layers of 512 and 4 neurons, another two candidates used the global convolution (g-conv) layers with 3 layers of 512, 512, and 4 kernels, or 2 layers of 512 and 4 kernels, and the rest 4 candidates adding a further global average pooling (GAP) layer in front of or behind an FC layer with 4 neurons or a g-conv layer with 4 kernels. Last, the candidate that yielded the best accuracy with the minimal size, number of parameters, and loss was equipped with three different normalizations, i.e., the local response normalization (LRN), batch normalization (BN), and group normalization (GN), further to improve the accuracy and training speed. The results show that the GAP layer followed by a single FC layer can best fuse the features passing down from the convolutional layers, compared with the multilayered FC or g-conv layers, significantly reducing the size and the number of parameters by over 89%, still with the improved accuracy. The highest accuracy was achieved at 98.02% out of the 8 fine-tuned models with the classification layer placed after GAP. Therefore, it was deemed as the GV. The GN, being not affected by the batch size, can make the validation converge more stably and further push the validation accuracy to 98.54%, when added after the convolutional modules. Since the learning rate has a great impact on the performance of the transfer learning model, it is necessary to choose a suitable initial value and a segmented constant decay strategy for the specific applications, which can effectively improve the model performance. The model proposed a novel idea for non-destructive monitoring of Ginkgo biloba seed germination and accurate prediction of germination rate after sowing.