Image Classification Method Based On Improved Lightweight Neural Network

With the continuous progress of science and technology and the continuous development of Internet technology,the number of pictures produced every day is growing geometrically.As the basis of computer vision tasks such as image recognition,image segmentation,image classification has been studied more widely.Traditional image classification methods focus on feature extraction and classifier selection.Most of the features extracted by the method are artificially designed,and there will be some deviation in the effect.In recent years,deep learning has achieved good results in the field of image classification,and has become the main research content of image classification.However,the image classification method based on deep learning has the shortcomings of memory intensive and computational intensive,which can not be better applied to low-memory mobile devices.The network model needs to be compressed and accelerated.This paper presents an improved model structure based on MobileNets,which can reduce parameters or improve classification accuracy by improving the MobileNets.1.An image classification method based on D-MobileNet model with local receptive field expansion is proposed.Because the Dilated Convolution core can expand the local receptive field without adding parameters,we use the Dilated Convolution core to replace the convolution core of a layer of the MobileNets model,expand the local receptive field of the layer,and extract a wider range of features.This method can improve the classification accuracy without increasing the number of parameters of the model.The experiments were carried out on Caltech-101,Caltech-256 and Uebingen Animals with Attributes data set.The results show that the D-MobileNet model can obtain about 2%higher classification accuracy than MobileNets.2.An image classification method based on densely connected B-MobileNet model is proposed.By introducing dense blocks into the MobileNets model,and through dense connections,we can make use of the feature map of the convolution layer in front of the dense blocks for many times,and reduce the parameters and calculation amount in the network model by setting appropriate superparametric growth rate.In this paper,the B-MobileNet with dense blocks is introduced to perform comparative experiments on Caltech-101 and Uebingen Animals with Attributes data sets.The results show that the parameters of B-MobileNet model can be reduced to 1/2-1/3 of the original network model,the speed can be increased by about two times,and the accuracy can be reduced by up to 1%.Among them,the classification accuracy of B2-MobileNet model on two datasets has also increased.