摘要
针对传统卷积神经网络在处理图像分类的问题过程中,需要较长的训练时间、大量的存储空间和计算资源等问题,提出一种基于卷积神经网络迭代优化的图像分类算法。对卷积神经网络进行初始化,在训练网络的过程中,对每层网络单个特征图的输入进行BN(batch normalize)处理,得到归一化的数据后再输入到特征图中,采用迭代的方法调整卷积神经网络中的参数并删除低于阀值的连接。实验结果表明,在Mnist、Cifar-10数据集上,经过迭代优化后的卷积神经网络分别提高了0.33%和3.42%的准确率,有效降低了网络中参数的数量,相比原始卷积神经网络中的参数数量分别降低87.94%、85.91%,网络收敛速度更快,减小了网络的训练时间。
Concerning the problem that it requires longer training time, plenty of storage and computing resources in traditional training of image classification task, an algorithm of image classification was proposed based on convolution neural network of iterative optimization. The convolution neural network was initialized and each layer of network in the process of training the net- work was batch normalized, the data were put into the feature map when finished processing. The parameter iterative method was used to adjust the convolutional neural network in connection and the connection below the threshold was deleted. Experimental results show that in Mnist and Cifar-10 data set, the convolutional neural network improves the accuracy by 0.33% and 3.42% respectively after the iterative optimization. It also reduces the number of parameters in the network effectively, com- pared to the original convolutional neural network, parameters are decreased by 85.91% and 87.94% respectively. Furthermore, the network convergence speed is higher and the network training time is reduced.
出处
《计算机工程与设计》
北大核心
2017年第1期198-202,214,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61373109
61003127)
关键词
卷积神经网络
特征图
网络连接
收敛
阈值
convolution neural network
feature map
connections of network
convergence
threshold
