VGG-S: Improved Small Sample Image Recognition Model Based on VGG16

Abstract

Convolutional Neural Network (CNN) has the problems of relying on large models, too long training time and over-relying on a large number of sample annotations. In this study, an improved image recognition model Vgg-Small (Vgg-S) based on Vgg16 is proposed. Based on the Vgg16 model, the Vgg16 model is pruned and improved to build a lightweight CNN model Vgg-S. Vgg-S can train with a small data set, and get better training results in a shorter training time. Through experiments on the public data set Caltech101, comparing common CNN prediction models, experiments prove that Vgg-S has a better performance on the small number of image recognition tasks.