基于自适应池化的双路卷积神经网络图像分类算法
|
摘要
针对相同构造的卷积神经网络输入同样的数据集也会提取到不同特征的情况,为利用该差异挖掘图像的深层特征,提出一种双路卷积神经网络模型的图像分类算法。在优化池化组合的基础上,在另一子网络中引入自适应池化丰富差异特征,提高特征表达层次;根据互补测量函数测量子网络间的特征差异的互补性,以此优化损失函数反向传播微调模型权重,提高图像分类的精准度。在MNIST和CIFAR-10图像集上的实验结果表明,基于自适应池化的双路卷积神经网络的分类能力优于现有的深度卷积神经网络。
Aiming at the situation that using the convolutional neural network with the same structured convolutional neural network inputting the same data set still extracts different features,to exploit this difference to mine the deep features of the image,a dual-channel convolution neural network model of image classification algorithm was presented.On the basis of optimizing the pooling combination,introducing adaptive pooling into another sub-network to enrich the characteristics of differences and improve the level of feature expression.The complementarity of feature differences was measured between sub-networks according to the complementary measurement function,so as to optimize the loss function,which fine-tuned the model weights by back propagation to improve the accuracy of image classification.Experimental results on the MNIST and CIFAR-10 image sets show that the dual-channel convolutional neural network based on adaptive convolution has better classification ability than the existing deep convolutional neural networks.
Aiming at the situation that using the convolutional neural network with the same structured convolutional neural network inputting the same data set still extracts different features,to exploit this difference to mine the deep features of the image,a dual-channel convolution neural network model of image classification algorithm was presented.On the basis of optimizing the pooling combination,introducing adaptive pooling into another sub-network to enrich the characteristics of differences and improve the level of feature expression.The complementarity of feature differences was measured between sub-networks according to the complementary measurement function,so as to optimize the loss function,which fine-tuned the model weights by back propagation to improve the accuracy of image classification.Experimental results on the MNIST and CIFAR-10 image sets show that the dual-channel convolutional neural network based on adaptive convolution has better classification ability than the existing deep convolutional neural networks.
引文
[1]Pang Y,Sun M,Jiang X,et al.Convolution in convolution for network in network[J].IEEE Transactions on Neural Networks&Learning Systems,2016,29(5):1587-1597.
[2]Cao J,Pang Y,Li X.Pedestrian detection inspired by appearance constancy and shape symmetry[J].IEEE Transactions on Image Processing,2015,26(10):2477-2486.
[3]YE Guolin,SUN Shaoyuan,GAO Kaijun,et al.Faster region convolution neural network based nighttime pedestrian detection[J].Laser&Optoelectronics Progress,2017,54(8):117-123(in Chinese).[叶国林,孙韶媛,高凯珺,等.基于加速区域卷积神经网络的夜间行人检测研究[J].激光与光电子学进展,2017,54(8):117-123.]
[4]MAO Ning,YANG Dedong,YANG Fucai,et al.Adaptive object tracking based on hierarchical convolution features[J].Laser&Optoelectronics Progress,2017,53(12):195-206(in Chinese).[毛宁,杨德东,杨福才,等.基于分层卷积特征的自适应目标跟踪[J].激光与光电子学进展,2017,53(12):195-206.]
[5]GAO Lin,WANG Junfeng,FAN Yong,et al.Robust visual tracking based on convolutional neural networks and conformal predictor[J].Acta Optica Sinica,2017,37(8):222-231(in Chinese).[高琳,王俊峰,范勇,等.基于卷积神经网络与一致性预测器的稳健视觉跟踪[J].光学学报,2017,37(8):222-231.]
[6]XU Lu,ZHAO Haitao,SUN Shaoyuan.Monocular infrared image depth estimation based on deep convolutional neural neworks[J].Acta Optica Sinica,2016(7):188-197(in Chinese).[许路,赵海涛,孙韶媛.基于深层卷积神经网络的单目红外图像深度估计[J].光学学报,2016(7):188-197.]
[7]ZHOU Junyu,ZHAO Yanming.Application of convolutional neural networks in image classification and target detection[J].Computer Engineering and Applications,2017,53(13):34-41(in Chinese).[周俊宇,赵艳明.卷积神经网络在图像分类和目标检测应用综述[J].计算机工程与应用,2017,53(13):34-41.]
[8]Akhlaqur R,Sumaria T.Ensemble classifiers and their applications:A review[J].International Journal of Computer Trends and Technology,2014,10(1):31-35.
[9]SONG Chao,XU Daoyun,QIN Yongbin.Separation of multiple convolutional neural networks[J].Computer Engineering,2017,43(6):145-149(in Chinese).[宋超,许道云,秦永彬.分离多路卷积神经网络研究[J].计算机工程,2017,43(6):145-149.]
[10]Zeiler MD,Fergus R.Stochastic pooling for regularization of deep convolutional neural networks[J].arXiv:1301.3557v1,2013.
[11]LIU Wanjun,LIANG Xuejian,QU Haicheng.Learning performance of convolutional neural networks with different pooling models[J].Journal of Image and Graphics,2016,21(9):1178-1190(in Chinese).[刘万军,梁雪剑,曲海成.不同池化模型的卷积神经网络学习性能研究[J].中国图像图形学报,2016,21(9):1178-1190.]
[12]Vedaldi A,Lenc K.MatConvNet-convolutional neural networks for MATLAB[C]//Proc of the ACM Int Conf on Multimedia,2015.
[2]Cao J,Pang Y,Li X.Pedestrian detection inspired by appearance constancy and shape symmetry[J].IEEE Transactions on Image Processing,2015,26(10):2477-2486.
[3]YE Guolin,SUN Shaoyuan,GAO Kaijun,et al.Faster region convolution neural network based nighttime pedestrian detection[J].Laser&Optoelectronics Progress,2017,54(8):117-123(in Chinese).[叶国林,孙韶媛,高凯珺,等.基于加速区域卷积神经网络的夜间行人检测研究[J].激光与光电子学进展,2017,54(8):117-123.]
[4]MAO Ning,YANG Dedong,YANG Fucai,et al.Adaptive object tracking based on hierarchical convolution features[J].Laser&Optoelectronics Progress,2017,53(12):195-206(in Chinese).[毛宁,杨德东,杨福才,等.基于分层卷积特征的自适应目标跟踪[J].激光与光电子学进展,2017,53(12):195-206.]
[5]GAO Lin,WANG Junfeng,FAN Yong,et al.Robust visual tracking based on convolutional neural networks and conformal predictor[J].Acta Optica Sinica,2017,37(8):222-231(in Chinese).[高琳,王俊峰,范勇,等.基于卷积神经网络与一致性预测器的稳健视觉跟踪[J].光学学报,2017,37(8):222-231.]
[6]XU Lu,ZHAO Haitao,SUN Shaoyuan.Monocular infrared image depth estimation based on deep convolutional neural neworks[J].Acta Optica Sinica,2016(7):188-197(in Chinese).[许路,赵海涛,孙韶媛.基于深层卷积神经网络的单目红外图像深度估计[J].光学学报,2016(7):188-197.]
[7]ZHOU Junyu,ZHAO Yanming.Application of convolutional neural networks in image classification and target detection[J].Computer Engineering and Applications,2017,53(13):34-41(in Chinese).[周俊宇,赵艳明.卷积神经网络在图像分类和目标检测应用综述[J].计算机工程与应用,2017,53(13):34-41.]
[8]Akhlaqur R,Sumaria T.Ensemble classifiers and their applications:A review[J].International Journal of Computer Trends and Technology,2014,10(1):31-35.
[9]SONG Chao,XU Daoyun,QIN Yongbin.Separation of multiple convolutional neural networks[J].Computer Engineering,2017,43(6):145-149(in Chinese).[宋超,许道云,秦永彬.分离多路卷积神经网络研究[J].计算机工程,2017,43(6):145-149.]
[10]Zeiler MD,Fergus R.Stochastic pooling for regularization of deep convolutional neural networks[J].arXiv:1301.3557v1,2013.
[11]LIU Wanjun,LIANG Xuejian,QU Haicheng.Learning performance of convolutional neural networks with different pooling models[J].Journal of Image and Graphics,2016,21(9):1178-1190(in Chinese).[刘万军,梁雪剑,曲海成.不同池化模型的卷积神经网络学习性能研究[J].中国图像图形学报,2016,21(9):1178-1190.]
[12]Vedaldi A,Lenc K.MatConvNet-convolutional neural networks for MATLAB[C]//Proc of the ACM Int Conf on Multimedia,2015.