Research On Deep Residual Networks Of Residual Networks For Image Classification

Image classification as one of the core technologies of computer vision has become a hot topic in artificial intelligence.The development of deep learning and convolutional neural network laid the technical foundation for image classification.To mitigate vanishing gradients and overfitting,this thesis proposes a method to optimize deep residual networks of residual networks(RoR)and realize accurate image classification by analyzing and studying the method of deep RoR.Firstly,RoR network is constructed by analyzing the residual network.Based on this,Pre-RoR and RoR-WRN networks are established to mitigate the problem of vanishing gradients and the RoR network is further optimized.Stochastic depth algorithm is used to effectively control network overfitting problem and save training time.Secondly,we put forward a optimization method of RoR by analyzing performance of three activation functions.We design different RoR architectures using three activation functions(ReLU,ELU and PELU),analyze the performance of image classification on different datasets,and propose the most effective optimization method.Thirdly,a Pyramidal RoR network model is proposed by analyzing the characteristics of RoR and PyramidNet.Then,we analyze the effect of different residual block structures on performance,and choose the residual block structure which best favour the classification performance.Finally,we add an important principle to further optimize Pyramidal RoR networks,drop-path is used to avoid overfitting and save training time.Image classification experiments are performed on CIFAR-10/100,SVHN and Adience datasets,and we achieve the current lowest classification error rates were 2.96%,16.40% and 1.59% on CIFAR-10/100 and SVHN respectively.And we achieve accuracy rate of 61.87% and 1-off accuracy rate of 93.39% on the Adience dataset.In this thesis,an optimized deep RoR networks is proposed to achieve accurate image classification.The best image classification results are obtained on CIFAR-10/100 and SVHN datasets.Experimental results show that the proposed method can effectively improve the accuracy of image classification.