Application Of Improved Capsule Network In Image Recognition

Image recognition can be widely applied in socio-economic.At present,deep learning has become the mainstream method of image recognition technology,such as Convolutional Neural Network(CNN),which is widely used.However,the recognition accuracy of CNN will be significantly reduced after the image changes such as rotation and translation,indicating that the spatial relationship of objects has a great impact on the recognition performance of the network.The emergence of the Capsule Network(Caps Net)has partially eliminated this effect.The network does not lose information about the location of objects.The translation,rotation and expansion of the object do not affect the accuracy of recognition.This makes the Caps Net model gradually replace the existing CNN model.However,the threshold of the loss function of this model is fixed,which is prone to premature "deactivation" capsules.Therefore,it is a meaningful topic to improve the Caps Net model.This paper improves the Caps Net from the following aspects:1)In the process of dynamic routing,spread loss is used to replace the original margin loss,whose function is to cancel the upper and lower limits of fixed threshold at the beginning of the model.The method of linear increase was used to change the threshold value to avoid premature "deactivation" capsule phenomenon to a certain extent.2)In order to verify the necessity of reconstruction,this paper compared the performance of capsule network with and without the reconstructed subnetwork under the same model parameters.The results show that the classification accuracy of capsule network model with reconstructed subnetwork is better than that of the latter.3)In order to determine the optimal parameters of the model,this paper studies different route iterations without adding reconstruction subnetwork,determines the influence of route iterations on classification accuracy,and then determines the optimal parameters of the model.The results show that when the number of routing iterations is 2 and the number of training rounds is 150,the classification accuracy of data set is the highest.Through data analysis,it is verified that the error rate of this model is reduced to 0.32% on MNIST dataset without enhancement and extension.The research shows that the improved capsule network improves the image recognition effect.At the same time,the improved capsule network also shows good performance in the Fashion-MNIST and CIFAR-10 datasets.