基于对称神经网络的图像分类

发布时间：2018-05-12 21:06

本文选题：差异特征 + 对称网络　；参考：《小型微型计算机系统》2017年03期

【摘要】：随着深度神经网络图像处理领域的研究与应用,图像分类精度得到了大幅提升.然而深度神经网络的随机处理过程导致同一网络对相同的训练图像,重复训练会提取到有差异性的特征.为了利用这种差异性,本文提出了一种对称神经网络模型,将两个特征维度相同的深度神经网络作为对称模型的左右子网络,通过前向传播得到有差异的图像特征,并在联合层对差异特征进行融合.为了优化网络,采用差异度量函数度量左右子网络的差异,用差异优化网络的损失函数,进而通过反向传播微调模型参数.基于上述思想,本文将扩展了左右子网络为深度置信网的对称深度置信网及左右子网络为卷积神经网络的对称卷积神经网络,在数据集MNIST和CIFAR-10上的实验测试表明,相较于深度置信网及卷积神经网络,该方式集成的对称深度模型能取得较好的分类性能.
[Abstract]:With the research and application of deep neural network image processing, the accuracy of image classification has been greatly improved. However, the random processing of the neural network leads to the same training image on the same network, and the repetition training will extract the characteristics of the difference. In order to use this difference, a symmetric neural network is proposed in this paper. The network model, the two characteristics of the same depth neural network as the symmetric model of the left and right sub networks, through the forward propagation of different image features, and the integration of the diversity of the characteristics of the fusion. In order to optimize the network, the difference measurement function is used to measure the difference between the right and left network, and the loss function of the network is optimized by difference. Based on the above thought, the symmetric convolution neural network of the left and right subnetworks is a symmetric depth confidence network of the deep confidence network and the left and right subnetworks are convolution neural networks. The experimental tests on the data set MNIST and CIFAR-10 show that the deep confidence network and the convolution neural network are compared. The symmetric depth model integrated in this way can achieve better classification performance.

【作者单位】：沈阳航空航天大学计算机学院;辽宁大学信息学院;
【基金】：国家自然科学基金项目(61170185)资助航空基金项目(2013ZC54011)资助辽宁省博士启动资金项目(20121034)资助辽宁省教育厅科学研究一般项目(L2014070)资助
【分类号】：TP391.41;TP183

【相似文献】