Improved Model Based on GoogLeNet and Residual Neural Network ResNet

Xuehua Huang

OAI: oai:igi-global.com:313442 • DOI: 10.4018/IJCINI.313442

To improve the accuracy of image classification, a kind of improved model is proposed. The shortcut is added to GoogLeNet inception v1 and several other ways of shortcut are given, and they are GRSN1_2, GRSN1_3, GRSN1_4. Among them, the information of the input layer is directly output to each subsequent layer in the form of shortcut. The new improved model has the advantages of multi-size and small convolution kernel in the same layer in the network and the advantages of shortcut to reduce information loss. Meanwhile, as the number of inception blocks increases, the number of channels is increased to deepen the extraction of information. The GRSN, GRSN1_2, GRSN1_3, GRSN1_4, GoogLeNet, and ResNet models were compared on cifar10, cifar100, and mnist datasets. The experimental results show that the proposed model has 3.07% improved to ResNet on data set cifar10, 2.08% on data set cifar100, 17.69% improved to GoogLeNet on data set cifar10, 28.47% on data set cifar100.