摘要
提出多视图卷积神经网络模型MV-PearlNet,替代人工进行细粒度珍珠分类.该模型采用并行化处理方式,针对珍珠的多个视角图片提取特征,可提升珍珠图片的特征提取效果,并且采用中间层特征融合作为珍珠的特征表达.在训练集数据量有限的情况下,通过MV-PearlNet结合K-means方法,将无监督聚类算法应用到提取得到的特征中,并利用相似度计算完成自动类标学习,这些操作起到了扩充数据集的作用,有助于改善深度分类模型因为训练集不足导致的欠拟合问题,可提高模型的分类准确率.实验结果表明,相比于主流卷积神经网络模型,MV-PearlNet对珍珠细粒度图片的分类准确率有明显的提高.
In this paper,a multi-view convolutional neural network model(MV-PearlNet)is proposed,which can replace artificial fine-grained pearl classification.The model uses parallel processing to extract features from multiple perspective pictures of pearls,which can improve the effect of feature extraction.In addition,MV-PearlNet model adopts the feature fusion of the middle layer as the feature expression of the pearl.When the amount of data in the training set is limited,MV-PearlNet is combined with K-means method in this paper.And,we apply the unsupervised clustering algorithm to the extracted features,when the similarity calculation is used to complete the automatic class-based learning.It is worth pointing out that the proposed MV-PearlNet model expands the data set and improves the under-fitting problem of the deep classification model due to insufficient training set.Therefore,it can improve the classification accuracy of the model.Experimental data indicates that MV-PearlNet has significantly improved the classification accuracy of pearl fine-grained pictures,compared with mainstream convolutional neural network models.
作者
钱涛
熊晖
陈晋音
QIAN Tao;XIONG Hui;CHEN Jin-yin(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第1期185-190,共6页
Journal of Chinese Computer Systems
基金
浙江省认知医疗工程技术中心开放基金项目(2018KFJJ07)资助
浙江省自然(LY19F020025)资助。
关键词
珍珠分类
多视图卷积神经网络
无监督聚类算法
主动类标学习
特征融合
pearl classification
multi-view
convolutional neural network
unsupervised clustering algorithm
automatic label learning
feature fusion
