摘要
为解决不同数据集共同类图像特征学习能力弱的问题,采用深度学习算法模型,提出一种基于多重卷积神经网络的跨数据集图像分类方法。以中值滤波预处理后的图像作为网络输入,在两个池化层之间采用两组连续卷积层,卷积特征提取和池化后,采用L2范数正则化的Softmax损失函数作为模型分类器,完成多重卷积神经网络分类的训练和测试。实验结果表明,相比于传统JDA方法、TCA方法和KPCA方法,该方法在经典数据集Caltech256、Amazon、Webcam和Dslr上具有更好的特征提取能力和更高的平均准确率。
To solve the problem of weak learning ability of common data in different datasets,a method of image classification based on multi-hconvolutional neural networks was proposed.The image was pre-processed with median filter as the network input.Two sets of continuous convolution layers were used between the two pool layers.After the convolution feature was extracted and pooled,the normalized Softmax loss function of L2 norm was used as a model classifier to complete the training and testing of multi-hconvolutional neural network classification.Experimental results show that compared with the traditional JDA method,TCA method and KPCA method,the proposed method has better feature extraction ability and higher average accuracy rate on classical datasets Caltech256,Amazon,Webcam and Dslr.
作者
刘鑫童
刘立波
张鹏
LIU Xin-tong;LIU Li-bo;ZHANG Peng(School of Information Engineering,Ningxia University,Yinchuan 750021,China)
出处
《计算机工程与设计》
北大核心
2018年第11期3549-3554,共6页
Computer Engineering and Design
基金
宁夏自然科学基金项目(NZ17010)
国家自然科学基金项目(61751215
61363054)
西部一流大学科研创新基金项目(ZKZD2017005)
关键词
跨数据集分类
卷积神经网络
多重卷积
特征学习
L2正则化
cross-dataset classification
convolution neural network
multi-convolutional
feature learning
L2 regularization
