图像分类卷积神经网络的反馈损失计算方法改进  

Improved Loss Calculation Algorithm for Convolutional Neural Networks in Image Classification Application

作　　者：周非[1,2] 李阳[1,2] 范馨月[1,2] ZHOU Fei;LI Yang;FAN Xin-yue(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Chongqing Key Laboratory of Optical Communication and Networks,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学光通信与网络重点实验室,重庆400065

出　　处：《小型微型计算机系统》2019年第7期1532-1537,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61471077)资助

摘　　要：当前在图像分类领域,卷积神经网络主要通过反向传播算法训练权重和偏置.在参数的训练过程中,网络的实际输出与样本标签之间的反馈损失计算方式会影响到卷积神经网络对图像的最终分类性能.本文研究发现,当增大训练样本标签的维度,提高不同类别标签间的最小汉明距离,并通过sigmoid激活函数结合交叉熵计算反馈损失时,所得到的卷积网络模型对图像的分类能力优于使用softmax激活函数结合独热编码计算反馈损失所得到的卷积网络模型的分类能力.本文使用多种卷积神经网络结构,并结合多个数据集进行训练和测试,所得到的仿真结果证明了本文观点的正确性.In recent years,convolutional neural networks(CNN)in image classification generally update the weights and offsets through the back-propagation algorithm.Therefore,a suitable algorithm to calculate the feed-back loss is pretty important for CNN,which greatly affects the classification accuracy of the CNN.In this paper,an improved loss calculation algorithm is proposed to improve the image classification accuracy of the CNN.The simulation experiments of this paper show that increasing the dimensions and hamming distance of training sample tags has a positive effect on the classification accuracy of CNN.And in terms of the image classification accuracy,the method of using the sigmoid activation function and the crossentropy loss function to calculate the feedback loss performs better than the method of using softmax activation function and the log-likelihood loss function to calculate the feedback loss.

关 键 词：图像分类 卷积神经网络 sigmoid激活函数 交叉熵损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]