The Improvement Of Robustness And Prediction Accuracy For Capsule Networks

Convolutional neural networks have become an important part of many computer vision tasks,such as image classification,object detection,and image generation,etc.However,compared with the traditional Convolutional Neural Networks,the recently emerged Capsule Neural Networks have the merit of capturing the spatial hierarchy between different components of an object and thus possess better classification accuracy for object recognition such as rotation,scaling and even overlapping.However,like traditional convolutional neural networks,capsule networks are vulnerable to adversarial attacks,resulting in unreliable predictions,which hinders the application of neural networks in critical areas such as medical and security.Therefore,studying the adversarial robustness and prediction accuracy of capsule networks will facilitate their application in real-life and production.In this paper,we study Capsule Networks from two aspects.On the one hand,we propose a regularization method based on the Lipschitz constant constraint to address the adversarial robustness of capsule networks,and use the adversarial training method to train a more robust Caps Net.Compared with other improved models,our method is computationally simple,has no structural changes to the capsule network,and retains the characteristics of the original model.Experimental results show that the improved model achieves improved robustness on MNIST and SVHN datasets,especially on the Fashion-MNIST dataset,where the prediction accuracy of the adversarial samples is improved up to 8% compared to the current model under strong attack algorithms.On the other hand,we investigate the problem of classification accuracy of capsule networks.We use the reconstruction module of the capsule network to obtain richer augmented samples by incorporating the contrast learning method into the training process and using the reconstructed samples as positive samples to improve the prediction accuracy of the network.Compared with existing methods,our model has a simple structure and better generalizability.The experimental results show that better prediction accuracy is achieved on MNIST and EMNIST datasets.