基于改进MobileNetV2的玉米籽粒图像品种的快速鉴别研究

Rapid Identification of Maize Seed Images Based on Improved MobileNetV2

为实现玉米籽粒品种的快速鉴别与保护,实验提出了基于改进MobileNetV2的玉米籽粒品种识别方法。采集了11种玉米籽粒图像共3938张,建立胚面与非胚面的双面混合数据集。按照7∶2∶1的比例随机划分训练集、验证集和测试集。对MobileNetV2网络模型进行微调改进,探讨全连接层数量与维度以及dropout的取值对模型性能的影响,并在此基础上解冻部分骨干网络,最终模型准确率达到0.979 5,相较于基准模型准确率(0.948 7)提高0.030 8。实验结果表明,迁移学习时对基准模型微调是十分有必要的,可以有效提高模型准确率。

To achieve fast identification and protection of maize seed varieties,in this paper,a maize seed variety recognition method based on a combination of migration learning and improved MobileNetV2 was proposed.Total of 3938 images of 11 maize seed varieties were collected to build a two-sided hybrid dataset of germ surface and non-germ surface.The training set,validation set and test set were randomly divided in accordance with the ratio of 7∶2∶1.The MobileNetV2 network model was fine-tuned and improved to explore the impact of the number and dimension of fully connected layers and the value of dropout on the performance of the model,and based on this,part of the backbone network was unfrozen,and the final model accuracy reached 0.979 5,an improvement of 0.030 8 compared to the benchmark model accuracy(0.948 7).The experimental results indicated that fine-tuning the baseline model was necessary for migration learning and could effectively improve the model accuracy.