The Research For Convolutional Nerual Networks On Image Classification

Convolution Neural Networks(ConvNets)is developed recent years,which is a method of pattern recognition combined deep learning and artificial neural network.Compared to the traditional methods for classification,ConvNets have a strong ability of learning the high dimensional features and have been widely applied in image classification,voice recognition,pedestrian detection and other fields.However,the development of ConvNets is facing many challenges.The large number of parameters of ConvNets result in problems such as slow response,gradient vanish and overfitting.Due to the large models,it is difficult to apply ConvNets on mobile devices which are lack of ability for computing and storing.To address the issues above,CDLN,which is a novel structure with multiple classifiers,is introduced to improve the response speed of ConvNets.A training method named Weighted Joint Training with Multiple Classifiers is proposed to optimize the performance of the network.Then we propose a method,OCS-HashNets,for compressing the network.In the experiments,multiple datasets is used to validate the algorithms.Specific contents and results of the research are as follow.1.A Weighted Joint Training with Multiple Classifiers,WJTMC,is proposed for CDLN.The multiple classifiers of CDLN can effectively improve the response speed of ConvNets.But the classification accuracy is not improved.Based on the shortcomings of CDLN,A method named JTMC is proposed firstly.The method is characterized by the fact that updating weights of a certain layer is related to the loss of all the classifiers behind when training the network.Thus,the samples for training increase and overfitting is prevented.As a result,the classification accuracy is improved.For further improving the classification performance,WJTMC is proposed based on JTMC which adds weight value to the loss of the classifiers.Experiments show WJTMC improves the classification performance of each classifier in the network so that the overall classification accuracy is enhanced.In addition,the response of the network does not decrease.2.OCS-HashNets is proposed to compressing the network.Firstly,we use the one-shot whole network compression scheme,OSC,to compress the entire network which can be divided into three steps.(1)Using the variational Bayesian matrix decomposition,VBMF,for rank selection.(2)Tucker decomposition for kernel tensor.(3)Fine tuning the network.Then we apply HashNets for weights sharing.This method randomly groups the connections weights into a number of hashing buckets and all connection weights within the same buckets share a single parameter value.It is worth noting that a low-cost hash function is used to group the weights so the method does increase the storage requirement of the network.In the experiments,the compressed network is performed on CPU and mobile device.Results show the model is greatly compressed and the speed and accuracy of the network are improved.