Research On Key Technologies Of Siamese Stacked Capsule Autoencoder

With the rapid development of machine learning,deep metric learning has gradually become a research hotspot.How to accurately measure the similarity between objects has always been one of the most critical problems in the field of deep metric learning.In the existing siamese network models that rely on pairs or triples,its similarity measurement ability is mainly limited by the ability to obtain image feature information.At present,most siamese network structures take convolution neural network as the backbone network.However,with the increasing accuracy requirements of similarity measurement tasks,some defects of convolution neural network are gradually exposed,such as the information loss of pooling operation in convolution.In2017,capsule network was proposed,showing better feature extraction ability than convolutional neural network.The neurons of the capsule network exist in the form of vectors.Each neuron can record more characteristic information than the traditional scalar neurons.Stacked capsule autoencoder was proposed in 2019,which is a neural network model that can learn features in an unsupervised way,and shows excellent feature extraction ability.In this thesis,a siamese network structure with excellent feature extraction performance of capsule network is designed by combining capsule network,stacked capsule autoencoding and siamese network.The specific research contents are as follows.1.The original capsule network is applied to complete the image recognition task with small picture size and low feature complexity.In order to promote the siamese capsule network to more practical applications,this thesis adjusts various structural parameters of the capsule network,so that the siamese capsule network can be applied to more scenes.2.The stacked capsule autoencoder can train the model in an unsupervised way,and has excellent feature extraction ability and strong affine transformation robustness.It provides a new possibility for network model training based on small samples.In this paper,a siamese network structure based on stacked capsule self encoder is designed,which adopts the training method of unsupervised training first and then fine-tuning.3.The siamese stacked capsule autoencoder is tested on CIFAR-10,CIFAR-100 and Fashion MNIST data sets respectively.It is found that the recognition accuracy has been steadily improved after applying the stacked capsule autoencoder to the siamese network.4.Aiming at the high standard of appearance patent image retrieval task in measurement accuracy and robustness of view transformation,capsule network happens to show the performance of traditional neural network.Therefore,a similarity measurement algorithm of appearance patent image based on siamese stacked capsule autoencoder is proposed.