It is essential to utilize deep-learning algorithms based on big data for the implementation of the new generation of artificial intelligence. Effective utilization of deep learning relies considerably on the number o...It is essential to utilize deep-learning algorithms based on big data for the implementation of the new generation of artificial intelligence. Effective utilization of deep learning relies considerably on the number of labeled samples, which restricts the application of deep learning in an environment with a small sample size. In this paper, we propose an approach based on a generative adversarial network (GAN) combined with a deep neural network (DNN). First, the original samples were divided into a training set and a test set. The GAN was trained with the training set to generate synthetic sample data, which enlarged the training set. Next, the DNN classifier was trained with the synthetic samples. Finally, the classifier was tested with the test set, and the effectiveness of the approach for multi-classification with a small sample size was validated by the indicators. As an empirical case, the approach was then applied to identify the stages of cancers with a small labeled sample size. The experimental results verified that the proposed approach achieved a greater accuracy than traditional methods. This research was an attempt to transform the classical statistical machine-learning classification method based on original samples into a deep-learning classification method based on data augmentation. The use of this approach will contribute to an expansion of application scenarios for the new generation of artificial intelligence based on deep learning, and to an increase in application effectiveness. This research is also expected to contribute to the comprehensive promotion of new-generation artificial intelligence.展开更多
高光谱成像技术在实现低成本大范围的土壤重金属快速监测方面独具潜力。针对高光谱图像反演中突出的小样本问题,本文基于分数阶微分FOD(Fractional Order Derivative)提出一种面向高光谱图像的土壤重金属反演方法。首先,利用土壤采样点...高光谱成像技术在实现低成本大范围的土壤重金属快速监测方面独具潜力。针对高光谱图像反演中突出的小样本问题,本文基于分数阶微分FOD(Fractional Order Derivative)提出一种面向高光谱图像的土壤重金属反演方法。首先,利用土壤采样点的邻近像元进行样本扩充,增加样本的光谱差异性;其次,采用FOD突出光谱特征同时保留微分光谱的渐变信息;进而通过竞争自适应重加权采样CARS(Competitive Adaptive Reweighted Sampling)优选波段,采用偏最小二乘方法(PLSR)建立反演模型。以新疆维吾尔自治区哈密市黄山南矿区获取的72个土壤样本和航空高光谱图像为研究数据,对铅(Pb)、锌(Zn)、镍(Ni)3种重金属进行反演,结果表明:样本扩充不仅缓和了模型的过拟合现象,还提升了重金属反演精度;最佳阶数的分数阶微分能有效增强光谱特征,提高反演精度;CARS相对于相关系数法CC(Correlation Coefficient)、遗传算法GA(Genetic Algorithm)选出的波段组合反演精度更优,对研究区重金属Pb、Zn、Ni的反演精度R2分别为0.7974、0.8690和0.8303,反演方法具有较好的鲁棒性。展开更多
构建大规模茶芽目标检测数据集是一项耗时且繁琐的任务,为了降低数据集构建成本,探索少量标注样本的算法尤为必要。本文提出了YSVD-Tea(YOLO singular value decomposition for tea bud detection)算法,通过将预训练模型中的基础卷积替...构建大规模茶芽目标检测数据集是一项耗时且繁琐的任务,为了降低数据集构建成本,探索少量标注样本的算法尤为必要。本文提出了YSVD-Tea(YOLO singular value decomposition for tea bud detection)算法,通过将预训练模型中的基础卷积替换为3个连续的矩阵结构,实现了对YOLOX算法结构的重构。通过维度变化和奇异值分解操作,将预训练权重转换为与重构算法结构相对应的权重,从而将需要进行迁移学习的权重和需要保留的权重分离开,实现保留预训练模型先验信息的目的。在3种不同数量的数据集上分别进行了训练和验证。在最小数量的1/3数据集上,YSVD-Tea算法相较于改进前的YOLOX算法,mAP提高20.3个百分点。对比测试集与训练集的性能指标,YSVD-Tea算法在测试集与训练集的mAP差距仅为21.9%,明显小于YOLOX的40.6%和Faster R-CNN的55.4%。在数量最大的数据集上,YOLOX算法精确率、召回率、F1值、mAP分别为86.4%、87.0%、86.7%和88.3%,相较于对比算法均最高。YSVD-Tea在保证良好性能的同时,能够更好地适应少量标注样本的茶芽目标检测任务。展开更多
基金the National Natural Science Foundation of China (91646102, L1724034, L16240452, L1524015, and 20905027)the MOE (Ministry of Education in China) Project of Humanities and Social Sciences (16JDGC011)+3 种基金the Chinese Academy of Engineering’s China Knowledge Center for Engineering Sciences and Technology Project (CKCEST-2018-1-13)the UK– China Industry Academia Partnership Programme (UK-CIAPP/260)Volvo-Supported Green Economy and Sustainable Development at Tsinghua University (20153000181)the Tsinghua Initiative Research Project (2016THZW).
文摘It is essential to utilize deep-learning algorithms based on big data for the implementation of the new generation of artificial intelligence. Effective utilization of deep learning relies considerably on the number of labeled samples, which restricts the application of deep learning in an environment with a small sample size. In this paper, we propose an approach based on a generative adversarial network (GAN) combined with a deep neural network (DNN). First, the original samples were divided into a training set and a test set. The GAN was trained with the training set to generate synthetic sample data, which enlarged the training set. Next, the DNN classifier was trained with the synthetic samples. Finally, the classifier was tested with the test set, and the effectiveness of the approach for multi-classification with a small sample size was validated by the indicators. As an empirical case, the approach was then applied to identify the stages of cancers with a small labeled sample size. The experimental results verified that the proposed approach achieved a greater accuracy than traditional methods. This research was an attempt to transform the classical statistical machine-learning classification method based on original samples into a deep-learning classification method based on data augmentation. The use of this approach will contribute to an expansion of application scenarios for the new generation of artificial intelligence based on deep learning, and to an increase in application effectiveness. This research is also expected to contribute to the comprehensive promotion of new-generation artificial intelligence.
文摘构建大规模茶芽目标检测数据集是一项耗时且繁琐的任务,为了降低数据集构建成本,探索少量标注样本的算法尤为必要。本文提出了YSVD-Tea(YOLO singular value decomposition for tea bud detection)算法,通过将预训练模型中的基础卷积替换为3个连续的矩阵结构,实现了对YOLOX算法结构的重构。通过维度变化和奇异值分解操作,将预训练权重转换为与重构算法结构相对应的权重,从而将需要进行迁移学习的权重和需要保留的权重分离开,实现保留预训练模型先验信息的目的。在3种不同数量的数据集上分别进行了训练和验证。在最小数量的1/3数据集上,YSVD-Tea算法相较于改进前的YOLOX算法,mAP提高20.3个百分点。对比测试集与训练集的性能指标,YSVD-Tea算法在测试集与训练集的mAP差距仅为21.9%,明显小于YOLOX的40.6%和Faster R-CNN的55.4%。在数量最大的数据集上,YOLOX算法精确率、召回率、F1值、mAP分别为86.4%、87.0%、86.7%和88.3%,相较于对比算法均最高。YSVD-Tea在保证良好性能的同时,能够更好地适应少量标注样本的茶芽目标检测任务。
