基于近红外光谱的家蚕种茧雌雄鉴别模型在多设备和多品种间的迁移

陈楚汉 1; 钟苏苑 2; 钟杨生 3; 王先燕 2; 代　芬 1

文章摘要

陈楚汉 1 ，钟苏苑 2，钟杨生 3，王先燕 2，代　芬 1.基于近红外光谱的家蚕种茧雌雄鉴别模型在多设备和多品种间的迁移[J].广东农业科学,2021,48(2):129-137

查看全文 HTML 基于近红外光谱的家蚕种茧雌雄鉴别模型在多设备和多品种间的迁移

Transfer of Sex Identification Model for Silkworm Cocoons Based on Near-infrared Spectrum Among Multiple Devices and Varieties

DOI：10.16768/j.issn.1004-874X.2021.02.017

中文关键词: 卷积神经网络迁移学习家蚕种茧近红外光谱雌雄检测

英文关键词: convolutional neural networks（CNN） transfer learning silkworm cocoon near-infrared spectrum sex identification

基金项目:国家自然科学基金（61675003）；广东省自然科学基金（2018A030310153）；广州市科技计划项目（201707010346）

作者	单位
陈楚汉 1 ，钟苏苑 2，钟杨生 3，王先燕 2，代　芬 1	1. 华南农业大学电子工程学院，广东广州 510642；2. 广东省蚕业技术推广中心，广东广州 510640； 3. 华南农业大学动物科学学院，广东广州 510642

摘要点击次数: 1698

全文下载次数: 623

中文摘要:

【目的】在家蚕种茧雌雄鉴别设备推广应用中，为了减少新品种、新设备需要重新建模的难度和时间，进行了基于近红外光谱的家蚕种茧雌雄鉴别模型在多设备和多品种间的迁移研究。【方法】首先使用 2 种型号光谱仪（NirQuest512 光谱仪和 SW2540 光谱仪）采集 4 个品种蚕茧（9 芙、7 湘、7 夏和 932）的近红外漫透射光谱数据；然后用源域数据集训练的卷积神经网络作为源域模型，并对其中的中间层输出进行可视化分析；针对不同的种茧品种和不同的采集设备对源域模型进行微调，构建了迁移后的模型；最后将迁移后的模型预测准确率与卷积神经网络（Convolutional neural networks，CNN）、支持向量机（Support vector machine，SVM）和随机森林（Random forests，RF）算法进行对比。【结果】采用源域 1 700 个 9 芙样本（NirQuest512 光谱仪）构建的 CNN 源域模型，具有很好的雌雄分辨能力，分辨准确率达到 99% 以上。以源域 CNN 模型中间层输出作为输入构建的 SVM 和 RF 模型，雌雄分辨准确率分别为 92% 和 90% 以上，通过可视化分析表明卷积层能很好地提取雌雄特征。对于目标域中样本数量较少的 100 个 7 湘（NirQuest512 光谱仪）、77 个 7 夏和 112 个 932 的样本（SW2540 光谱仪），训练集比例为 70% 时，通过微调源域 CNN 模型后得到的目标域 CNN 模型的准确率分别为 96.90%、 99.67%、97.29%，效果最优；独立的 SVM 模型准确率分别为 92.49%、94.25%、93.65%，效果次之；独立 RF 模型的准确率分别为 80.93%、80.17%、81.47%，效果稍差；独立 CNN 模型的准确率分别只有 58.87%、56.33%、 72.17%，效果最差。通过多次不同训练集数量的建模比较，同样显示在数据量较少的情况下，迁移学习后 CNN 模型最优，传统机器学习方法次之，深度 CNN 模型最差。【结论】不同光谱仪或者不同品种的情况下深度迁移学习模型的可迁移性，为使用多种光谱仪和采集多种品种蚕茧时快速建立一个蚕茧雌雄分类模型提供了理论和实践的依据。

英文摘要:

【Objective】In the promotion and application of sex identification equipment of silkworm cocoons, in order to reduce the difficulty and time of new varieties and new equipment to be re-modeled, the transfer of a sex identification model for silkworm cocoons based on near-infrared spectrum among multiple devices and varieties was studied.【Method】Firstly, two kinds of spectrometers（NirQuest512 spectrometer and SW2540 spectrometer）were used to collect the near-infrared diffuse transmission spectrum data of four silkworm varieties（9fu, 7xiang, 7xia and 932）, then the convolutional neural networks trained by source domain dataset were used as the source domain model, and the outputs of the middle layer were analyzed visually. Furthly, the source domain model was tuned finely according to different silkworm varieties and different collection devices, and the transferred model was constructed. Finally, the model prediction accuracy after transfer learning was compared with the algorithm of Convolutional Neural Networks（CNN）, Support Vector Machine（SVM）and Random Forests （RF）. 【Result】The results show that the CNN source domain model constructed with 1 700 samples of 9fu（NirQuest512 spectrometer）has good sex resolution ability, and the accuracy of sex resolution is over 99%. The SVM and RF models were constructed with the outputs of middle layer of CNN source domain model as inputs, and the accuracy of sex discrimination is over 92% and 90%, respectively. Visual analysis shows that the convolutional layer can well extract the sex characteristics. For the target domain with less sample size of 100 7xiang（NirQuest512 spectrometer）, 77 7xia and 112 932（SW2540 spectrometer）, when the proportion of the training set is 70%, by fine-tuning the source domain CNN model, the accuracy of the obtained target domain CNN model is 96.90%, 99.67% and 97.29% respectively, with the optimal effect; the accuracy of the independent SVM model is 92.49%, 94.25% and 93.65% respectively, with the effect slightly lower than that of CNN; the accuracy of the independent RF model is 80.93%, 80.17% and 81.47% respectively, with the effect worse than that of SVM; the accuracy of the independent CNN model is only 58.87%, 56.33% and 72.17% respectively, with the worst effect. Through modeling comparison of the number of different training sets for many times, it is also shown that, in the case of less data, the effect of CNN model after transfer learning is the best, followed by the traditional machine learning method, and that of the deep CNN model is the worst.【Conclusion】The above results show the mobility of the deep transfer learning model with different spectrometers or different species, which provides theoretical and practical basis for the rapid establishment of a sex classification model for silkworm cocoons with multiple spectrometers and various varieties of cocoons.

查看/发表评论下载PDF阅读器