Research On Image Clustering Based On Unsupervised Deep Learning

Today,data clustering is a basic problem in many fields,such as machine learning,pattern recognition,computer vision,data compression,etc.Image clustering,as a key technology in the image field,has very important research value and application prospects.With the rapid development of artificial intelligence technology,deep learning plays an important role in related fields such as images,so this thesis mainly studies the application of embedded clustering problems in unsupervised deep learning,and proposes a deep self-reliance based on Sliced-Wasserstein distance Encoding embedded clustering algorithm;secondly,a deep embedded clustering algorithm based on self-supervised Pretext task and noise contrast loss is proposed.The specific two kinds of algorithm innovation ideas and main work are as follows:Deep self-coding embedded clustering algorithm based on Sliced-Wasserstein distance.Because conventional autoencoders cannot constrain the latent variable feature space and result in poor clustering performance,this thesis applies the Sliced-Wasserstein distance to the clustering self-encoding network framework,which not only maintains the advantage of the Wasserstein distance,but also can pass the Sliced-Wasserstein The distance constrained latent variable feature space uses soft distribution to implement clustering in a limited space to optimize the performance of unsupervised embedded clustering of images.The experimental results on the MNIST,Fashion MNIST and USPS datasets show that the algorithm proposed in this thesis has improved the unsupervised clustering accuracy(ACC)and standardized mutual information(NMI)indicators,especially compared with the traditional clustering algorithm The improvement is obvious.Deep embedded clustering algorithm based on self-supervised Pretext task and noise contrast loss.Since the unsupervised clustering algorithm cannot use labels,it is difficult to ensure that the self-encoder can fully express the original information.Therefore,the unsupervised Pretext task algorithm is used to improve the unsupervised semantic information acquisition and obtain better feature generalization for the target task ability;redesigned the noise contrast loss function used to assist cluster optimization,further improving the performance index of clustering.The experimental results on the MNIST,Fashion MNIST and USPS datasets show that the algorithm proposed in this thesis has slightly improved the clustering indicators ACC and NMI,and the clustering effect is more obvious.