Research On Image Classification Algorithms Based On Convolutional Neural Networks

Image classification is of great significance for human society and daily life.Convolution neural networks,a new way of deep learning,have come to fruition in many fields of computer vision with the rapid development of hardware computing power.Convolutional neural networks are used in almost all aspects of life because of their efficiency – local connectivity,parameter sharing,and good scalability and robustness.Image classification tasks are among them.Although convolution neural networks are powerful,there are still many problems which impede the possibility for them to be better.This paper focus on the research and improvement of the topological structure of convolution neural networks.Residual network is a very successful convolutional neural network.There is a lot of research,even some non-computer vision research,use this architecture as their foundation to solve specific problems.However,it's not perfect.The down-sample part,the projection structure,doesn't satisfy the residual learning idea advocated by the residual network itself.The use of projection structure will lead to the inefficiency of the gradient flow in the network and the decrease of learning efficiency of the whole network.In this paper,Identity structure and No Projection structure is proposed,and its performance is verified on the CIFAR10/100 dataset.Experimental results show that these two structures proposed in this paper is able to boost the learning efficiency of the network.Under certain conditions,they can also improve the performance of the network.This paper also makes discussion about the improvement to the topology of the residual network.Densely connected network is also a very inspiring network.Its emphasis on the information flow in the network makes itself a great success in the field of computer vision.This paper proposes the idea that utilizing one-dimension convolution to increase the diversity of the information in the network.Experimental results show that the structure proposed in this paper can help improve the parameter efficiency and computational efficiency of the original densely connected network to a certain extent,making it more competitive.