基于局部感受野扩张D-MobileNet模型的图像分类方法

Image classification method based on D-MobileNet model

针对轻量级深度神经网络MobileNet会减少分类准确率的问题,将空洞卷积核引入MobileNet模型的某一卷积层中,提出一种基于局部感受野扩张的D-MobileNet模型。该模型根据空洞卷积核所在位置的不同分为三种结构,在不增加参数数量的同时能够扩大该层卷积核的局部感受野,提高分类精度。实验在Caltech-101数据集、Caltech-256数据集以及图宾根大学动物分类数据库上进行,结果表明,D-MobileNet模型可获得比MobileNet更好的分类准确率,最多可以提高2%。

Aiming at the problem that lightweight deep neural network MobileNet can reduce classification accuracy,this paper proposed a D-MobileNet(dilated convolution MobileNet)model based on local receptive field expansion by introducing dilated convolution kernel into a convolution layer of MobileNet model.The models consisted of three structures according to the location of the dilated convolution kernel.Without increasing the number of parameters,they could expand the local receptive field of the layer convolution kernel and improve the classification accuracy.This paper carried out the experiments on Caltech-101 database,Caltech-256 database and Uebingen animals with attributes database.The results show that the D-MobileNet models can achieve better classification accuracy than that of MobileNet,and can improve the classification accuracy by up to 2%.