基于改进的卷积神经网络的图像分类性能  

Research on Image Classification Performance Based on Improved Convolution Neural Network

作　　者：常祥[1] 杨明[2] CHANG Xiang a YANG Ming b(a. National Key Lab for Electronic Measurement and Technolog b. College of Science, North University of China, Taiyuan 030051, Chin)

机构地区：[1]中北大学信息探测与处理山西省重点实验室,太原030051 [2]中北大学理学院,太原030051

出　　处：《重庆理工大学学报（自然科学）》2017年第3期110-115,共6页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金资助项目(61171179)

摘　　要：将改进的卷积神经网络应用到图片目标识别中。为了提高分类预测准确度,对传统卷积神经网络结构进行了改进,其具体结构为:卷积层C1—池化层S1—卷积层C2—池化层S2—卷积层C3—池化层S3—全连接层FC—输出,主要增加了卷积层和池化层层数,且在卷积滤波器规格选择上统一选择了5×5。最后用这一网络结构模型和其他模型(Re Net、APAC、PACNet)对CIFAR-10数据库进行试验对比,通过最终的预测准确度可以看出:改进后的卷积神经网络的精度达90.37%,高于其他3种模型。An improved convolution neural network is applied to image object recognition. In order to improve the accuracy of classification prediction,this paper improves the structure of the basic convolution neural network. The concrete structure is as follows： Convolution layer C1—Pool layer S1—Convolution layer C2—Pool layer S2—Convolution layer C3—pool Layer S3—full-connect layer FC—output,ant it mainly increased the number of convolution layers,and unified selection of 5 × 5in the convolution filter specification. Finally,this model is compared with other models（ Re Net,APAC,PACNet） for CIFAR-10 database. Through the final prediction accuracy,it can be seen that the improved convolutional nerve has a better precision of 90. 37% than the other three models.

关 键 词：卷积神经网络 图像分类技术 卷积层 池化层 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] O175.1[自动化与计算机技术—计算机科学与技术]