Improved Capsule Network And Its Application In Deep Fake Detection

The capsule network transmits information in the form of vectors,which can retain more image information such as spatial location,and can represent the image with the corresponding intensity of the feature.A dynamic routing algorithm is used between layers to replace the pooling in traditional convolutional neural networks.This paper conducts research from the aspects of network model and dynamic routing algorithm,combined with the excellent technology in convolutional neural network,to further improve the performance of the capsule network,the specific research content is as follows:The use of a consistent parameter network reduces the computational complexity of the capsule network,while improving the dynamic routing algorithm and reducing overfitting.By keeping the routing coefficients of different capsules representing the same location consistent,the parameters to be trained in the dynamic routing process are reduced to 1/32 of the original.The reconstruction network was rebuilt using a convolutional decoder,and the network parameters were further reduced.Experiments on the MNIST data set can prove that the model proposed in this paper can reduce the amount of parameters and training time while ensuring the accuracy of the test,and the reconstructed image noise is smoother,which is better than the reconstructed image of the baseline capsule network.Aiming at the situation that the capsule network is not highly accurate on color data sets with complex backgrounds such as CIFAR-10,this paper proposes an improved capsule network using dense connections.In the feature extraction part of the network,the features of different dimensions are extracted and connected in a densely connected manner as the input of the capsule layer,at the same time,a Batch Normalization optimization method was added to alleviate the problem of gradient disappearance.The experiments on the CIFAR-10 dataset prove that the improved capsule network proposed in this paper has 16% more recognition accuracy than the baseline capsule network,and the reconstructed image has only a small amount of noise,which is more clear than the baseline capsule network.Finally,the improved capsule network is applied to deep forgery detection.Experiments on Forensics++ data show that the capsule network is more accurate than other methods.At the same time,the effects of different numbers of capsule networks on the results are tested.The experiment proves that capsule networks are more flexible in network structure than general deep learning networks.This article has improved the feature extraction,dynamic routing,and reconstruction network of the capsule network.The effectiveness of the improvement is verified through experiments,which improves the performance of the capsule network and provides a reference direction for future research and development.