为准确实时跟踪羊只目标,进行疾病异常预警,实现奶山羊精细化养殖,本文基于DiMP跟踪模型,利用奶山羊跟踪对象单一且图像样本丰富的特点,结合迁移学习和类特定融合方法,设计了一种类特定的奶山羊目标跟踪模型,能够有效克服DiMP算法在跟...为准确实时跟踪羊只目标,进行疾病异常预警,实现奶山羊精细化养殖,本文基于DiMP跟踪模型,利用奶山羊跟踪对象单一且图像样本丰富的特点,结合迁移学习和类特定融合方法,设计了一种类特定的奶山羊目标跟踪模型,能够有效克服DiMP算法在跟踪类特定目标时定位精度不足的缺点。利用构建的奶山羊视频跟踪数据训练集对跟踪算法进行迁移训练,加快模型收敛速度,使评估网络预测出的边界框更贴合奶山羊真实框的位置和尺寸。在线跟踪阶段,针对目标模板仅采用第1帧特征制作整个序列的调制向量,导致该调制向量相对整个跟踪阶段特征不具代表性,与后续帧差异大的缺点,使用训练集制作包含奶山羊各种姿态的类调制向量,以指数消融方式更新奶山羊类调制向量与第1帧调制向量间的比重,增强边界框回归任务中的奶山羊特征与背景的判别性。提出的算法在测试集上的AUC(Area under curve)和精准度(Precision)分别为76.20%和60.19%,比DiMP方法分别提升6.17、14.18个百分点,跟踪速度为30 f/s,满足实时跟踪的要求。实验结果表明,提出的类特定奶山羊目标跟踪方法可用于监测复杂场景下奶山羊的运动,为奶山羊精细化管理提供了技术支持。展开更多
在文本分类中,选取一个高效的分类算法是提高文本分类准确度,缩短分类时间的关键。提出基于指数分布族的多项式贝叶斯类特定分类算法(exponential family-multinomial naive Bayes,EF-MNB),基于多项式模型构造了 N 个类的分布,利用类特...在文本分类中,选取一个高效的分类算法是提高文本分类准确度,缩短分类时间的关键。提出基于指数分布族的多项式贝叶斯类特定分类算法(exponential family-multinomial naive Bayes,EF-MNB),基于多项式模型构造了 N 个类的分布,利用类特定特征选择算法得到第 N 个类的特征子集及对应类的特征概率密度函数(probability density function,PDF),通过指数分布族构造了 N 个类的原始PDF估计表达式,给定 N 个类的训练集,得到了第 N 个类的最优PDF估计,并基于贝叶斯定理制定了分类规则。仿真结果表明,与基于文档主题生成模型和支持向量机(latent dirichlet allocation-support vector machine,LDA-SVM)的层次分析分类算法、改进的超球支持向量机(improved hyper-sphere support vector machine,IHS-SVM)文本分类算法和基于主成份分析和k最近邻(principal component analysis-k-nearest-neighbor,PCA-KNN)混合分类算法相比,EF-MNB类特定分类算法使用少量的时间就可获得更高分类准确率。展开更多
Classification of plume and spatter images was studied to evaluate the welding stability. A high-speed camera was used to capture the instantaneous images of plume and spatters during high power disk laser welding. Ch...Classification of plume and spatter images was studied to evaluate the welding stability. A high-speed camera was used to capture the instantaneous images of plume and spatters during high power disk laser welding. Characteristic parameters such as the area and number of spatters, the average grayscale of a spatter image, the entropy of a spatter grayscale image, the coordinate ratio of the plume centroid and the welding point, the polar coordinates of the plume centroid were defined and extracted. Karhunen-Loeve transform method was used to change the seven characteristics into three primary characteristics to reduce the dimensions. Also, K-nearest neighbor method was used to classify the plume and spatter images into two categories such as good and poor welding quality. The results show that plume and spatter have a close relationship with the welding stability, and two categories could be recognized effectively using K-nearest neighbor method based on Karhunen-Loeve transform.展开更多
文摘为准确实时跟踪羊只目标,进行疾病异常预警,实现奶山羊精细化养殖,本文基于DiMP跟踪模型,利用奶山羊跟踪对象单一且图像样本丰富的特点,结合迁移学习和类特定融合方法,设计了一种类特定的奶山羊目标跟踪模型,能够有效克服DiMP算法在跟踪类特定目标时定位精度不足的缺点。利用构建的奶山羊视频跟踪数据训练集对跟踪算法进行迁移训练,加快模型收敛速度,使评估网络预测出的边界框更贴合奶山羊真实框的位置和尺寸。在线跟踪阶段,针对目标模板仅采用第1帧特征制作整个序列的调制向量,导致该调制向量相对整个跟踪阶段特征不具代表性,与后续帧差异大的缺点,使用训练集制作包含奶山羊各种姿态的类调制向量,以指数消融方式更新奶山羊类调制向量与第1帧调制向量间的比重,增强边界框回归任务中的奶山羊特征与背景的判别性。提出的算法在测试集上的AUC(Area under curve)和精准度(Precision)分别为76.20%和60.19%,比DiMP方法分别提升6.17、14.18个百分点,跟踪速度为30 f/s,满足实时跟踪的要求。实验结果表明,提出的类特定奶山羊目标跟踪方法可用于监测复杂场景下奶山羊的运动,为奶山羊精细化管理提供了技术支持。
文摘在文本分类中,选取一个高效的分类算法是提高文本分类准确度,缩短分类时间的关键。提出基于指数分布族的多项式贝叶斯类特定分类算法(exponential family-multinomial naive Bayes,EF-MNB),基于多项式模型构造了 N 个类的分布,利用类特定特征选择算法得到第 N 个类的特征子集及对应类的特征概率密度函数(probability density function,PDF),通过指数分布族构造了 N 个类的原始PDF估计表达式,给定 N 个类的训练集,得到了第 N 个类的最优PDF估计,并基于贝叶斯定理制定了分类规则。仿真结果表明,与基于文档主题生成模型和支持向量机(latent dirichlet allocation-support vector machine,LDA-SVM)的层次分析分类算法、改进的超球支持向量机(improved hyper-sphere support vector machine,IHS-SVM)文本分类算法和基于主成份分析和k最近邻(principal component analysis-k-nearest-neighbor,PCA-KNN)混合分类算法相比,EF-MNB类特定分类算法使用少量的时间就可获得更高分类准确率。
基金Project (51175095) supported by the National Natural Science Foundation of ChinaProjects (10251009001000001,9151009001000020) supported by the Natural Science Foundation of Guangdong Province,ChinaProject (20104420110001) supported by the Specialized Research Fund for the Doctoral Program of Higher Education of China
文摘Classification of plume and spatter images was studied to evaluate the welding stability. A high-speed camera was used to capture the instantaneous images of plume and spatters during high power disk laser welding. Characteristic parameters such as the area and number of spatters, the average grayscale of a spatter image, the entropy of a spatter grayscale image, the coordinate ratio of the plume centroid and the welding point, the polar coordinates of the plume centroid were defined and extracted. Karhunen-Loeve transform method was used to change the seven characteristics into three primary characteristics to reduce the dimensions. Also, K-nearest neighbor method was used to classify the plume and spatter images into two categories such as good and poor welding quality. The results show that plume and spatter have a close relationship with the welding stability, and two categories could be recognized effectively using K-nearest neighbor method based on Karhunen-Loeve transform.
