| In recent years,the unsupervised learning method in deep learning has attracted wide attention.As a deep learning network model of unsupervised learning,the Stacked Capsule Autoencoder extracts the components and attributes of the image in the pixel two-dimensional space by pixel reconstruction,and describes the relationship between the components and the whole in a reconstructed way.By improving and optimizing the Stacked Capsule Autoencoder,the feature extraction ability and classification accuracy of the model are further improved.The main research work and conclusions are as follows:(1)Optimization algorithm of Stacked Capsule Autoencoder based on manifold regularization.Aiming at the problem that the Stacked Capsule Autoencoder has slow detection performance and can’t better mine the local features of the image,the Scharr filter is introduced into the Stacked Capsule Autoencoder model to realize image reconstruction and improve the accuracy of image target detection,and the manifold regularization term is introduced into the loss function to strengthen the extraction of local features in the original data space.Finally,the Stacked Capsule Autoencoder based on manifold regularization is used to learn parameters and select more distinctive features.Experiments on MNIST and Fashion MNIST datasets show that compared with the original network structure,the optimization algorithm improves the image classification accuracy by 0.26% and 9.23% respectively,and the training speed of the model is also obviously improved.(2)Optimization algorithm of Stacked Capsule Autoencoder based on support vector machine.In order to solve the problem that K-MEANS has poor classification effect on image data sets with complex shape features and can not better classify the image features extracted by Stacked Capsule Autoencoder algorithm,the model is further improved on the basis of the optimization algorithm of Stacked Capsule Autoencoder based on manifold regularization.According to different coding types,Linear autoencoder,Convolutional autoencoder and Convolutional autoencoder based on self-attention mechanism are adopted and compared on the component capsule autoencoder,and the excellent coding type is determined.Support vector machines based on different kernel functions are used to classify image datasets,and more accurate classification results are obtained by comparing different kernel functions.Experiments with MNIST and Fashion MNIST datasets with different noise types show that the classification accuracy of the model is improved by 0.99% and 20.26% respectively,and better classification accuracy is obtained. |
