Research On Image Recognition Algorithm Based On Optimized Capsule Network

Artificial intelligence has become more and more close to people's daily,recognition of license plates,and payment operations through facial recognition h ave been integrated into all walks of life in society.Most of the core algorithms of these recognition systems come from convolutional neural networks(CNN).Although convolutional neural networks are already very mature and well applied algorithms in the field of image recognition,convolutional neural networks still exist.Problems,such as the inexplicability of convolutional neural networks,and the excessive emphasis on learning invariant features during the learning process,are all problems that convolutional neural networks cannot avoid.The capsule network has improved to these problems in the network structure.In this paper,the convolutional neural network and the capsule network are optimized.The stacked capsule autoencoder is used on the multiple data sets with the convolutional neural network and the capsule network.Compared with the experiment,through the experimental results and analysis,it is shown that the optimization method for convolutional neural network and capsule network proposed in this paper is effective.The main work of this article includes:(1)Detailed analysis of the research and application status of convolutional neural networks at home and abroad,an analysis of image recognition algorithms based on convolutional neural networks,and for its existing problems,the research ideas and research contents of this article are proposed.(2)Elaborated the N versions of the capsule network from 2017 to 2020,and used the latest stacked capsule autoencoder to complete the experimental content of this article.(3)Designed a comparative experiment between convolutional neural network and capsule network on MINST dataset.First,a comparative experiment is conducted in the case where the training set is one-to-one identification and the test set is also one-to-one identification.The experimental results show that the convolutional neural network and the capsule network perform similarly in this case.Then conduct a comparative experiment in the case where the training set is one-to-one identification and the test set is two-to-two identification.(4)Based on the experimental content of the previous two chapters,the structure of the capsule network was optimized,and the experiment was conducted again on the MNIST dataset.It was found that the optimized capsule network had higher training efficiency and a certain improvement in recognition accuracy.In addition,a hybrid batch training method was used to train the capsule network,and a better recognition accuracy was obtained.(5)The optimized capsule network was tested on the CIFAR-10,CIFAR-100,and SVHN data sets respectively,and it was found that the recognition accuracy of the capsule network can be slightly higher than that of other neural network algorithms on these data sets.