基于卷积神经网络的宠物猫品种分类研究  

Research on pet cat breed classification based on convolutional neural network

作　　者：王建霞[1] 张成 闫双双 WANG Jianxia;ZHANG Cheng;YAN Shuangshuang(College of Information Science and Engineering, Hebei University of Science and Technology, Shijiazhuang, Hebei 050018, China)

机构地区：[1]河北科技大学信息科学与工程学院,河北石家庄050018

出　　处：《河北工业科技》2020年第6期407-412,共6页Hebei Journal of Industrial Science and Technology

基　　金：河北省自然科学基金(F2018208116)。

摘　　要：为了提高宠物猫品种分类的准确率,提出了一种卷积神经网络融合的方法进行特征提取。首先,基于堆叠卷积自动编码器的域自适应技术,采用反卷积操作丰富特征图;其次,利用Inception结构增加网络的宽度来提取多尺度信息的特征图;最后,使用Softmox函数对图像进行分类,在Oxford-ⅢT数据集中进行实验分析。实验结果表明,利用改进后的模型对宠物猫进行分类,准确率高于对比模型,达到了84.56%,损失值为0.0150。所提出的卷积神经网络融合方法不仅能通过丰富特征图、加深网络深度更好地表达特征,还能提高分类性能和收敛性能,较好地解决了宠物品种识别中由宠物相似所带来的识别率低的问题,还可以推广应用到其他图像相似问题的应用场景中。In order to improve the accuracy of pet cat breed classification,a convolutional neural network fusion method was proposed for feature extraction.Firstly,based on the domain adaptive technology of the stacked convolutional autoencoder,the deconvolution operation was used to enrich the feature map;Secondly,the Inception structure was used to increase the width of the network to extract the feature map of multi-scale information;Finally,the images were classified by the Softmox function and were experimentally analyzed in the Oxford-ⅢT data set.The experimental results show that the accuracy of classifying pet cats by using the improved model is higher than that of the comparison model,reaching 84.56%,and the loss value is 0.0150.The proposed convolutional neural network fusion method can not only enrich feature maps and deepen the network depth to better express features,but also improve the classification performance and convergence performance.The method can better solve the problem of low recognition rate caused by pet similarity in pet breed recognition,and can also be extended to the application scenarios of other image similarity problems.

关 键 词：计算机图像处理 深度学习 卷积神经网络 反卷积 宠物猫分类 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]