Multi-label Image Classification And Semantic Segmentation Based On Capsule Networks

Since the concept of the capsule network was proposed,it has received widespread attention and use.The theoretical basis of the capsule network is that the human brain operates a mechanism,which passes low-level visual information to the neuron that it thinks can handle this information best.In addition to storing the object's characteristic information,the capsule in the machine can also store the pose information(position,orientation and angle)of the object,and the connection between the capsules can transfer the information without loss.Capsule network has a better performance in single-label image classification.Its classification results on MNIST and Cifar-10 datasets are better than convolutional neural networks.However,whether the capsule network can also perform well on more complex data sets or other areas of computer vision is still unknown.This paper analyzes these application limitations of the capsule network.Besides it combines the characteristics of the convolutional neural network for multi-label classification tasks and semantic segmentation to improve the structure of the capsule network so that it can be applied to multi-label classification tasks and semantic segmentation tasks.What's more,its performance can be better than the convolutional neural network.The main research work and contributions of this paper are summarized as follows:1.A capsule dynamic routing algorithm which is suitable for multi-label classification is proposed in this paper.By dispersing the activation amount of lowlevel capsules to high-level capsules,the capsule network can respond to multiple labels.At the same time,for the case where the input of the multi-label classification task is large,the paper put forward the idea that using the way of implanting the capsule layer into the convolution layer to eliminate the limitation of insufficient capacity of the capsule layer;and it is proposed to use probability parameters to bind the connections between capsules,which can reduce the strong coupling between capsules and alleviate the overfitting problem of the network.2.The capsule layer is used to imitate U-Net's symmetrical codec structure to design a capsule network for semantic segmentation.A local dynamic routing algorithm for the capsule network is proposed to reduce the number of connections between capsule layers and decrease the amount of parameters of the semantic segmentation capsule network.Aiming at the problem that the semantic segmentation network cannot segment the edges well,a general edge preservation algorithm is proposed.By constructing a branch network using the edge preservation algorithm,the network's attention to the target edge information can be enhanced and the segmentation effect can be improved.In conclusion,by improving the structure and dynamic routing mechanism of the original capsule network,this paper proposes a capsule network algorithm that is suitable for applications in multi-label image classification and semantic segmentation.Experiments show that the algorithm proposed in this paper performs better than convolutional neural networks on multiple data sets,which paves the way for future expansion and application of capsule networks.It is sure to have good application prospects and scientific research value.