Robust Subspace Clustering Algorithm Based On Deep Learning

With the arrival of the age of big data,the demand for data processing is increasing,and the application of subspace clustering is becoming more and more extensive,including the application in image video segmentation,outlier detection,face recognition,salient target detection and other fields,which also puts forward higher requirements for the progress of the algorithm.At present,the subspace clustering algorithm is lack in robustness of mining algorithm to data itself,and seldom consider to remove the noise of the data itself and modify the damaged data in different degrees to improve the performance of the algorithm.Therefore,the algorithm is difficult to meet the needs of complex data processing in the actual scene.In view of the above disadvantages,this paper draws on the knowledge of sparse representation and the advantages of u-net network in the field of noise reduction,innovatively proposes a Lu Bang subspace clustering algorithm based on deep learning,and makes a series of improvements.The main research and innovation points of this paper are as follows:1.A sparse robust subspace clustering algorithm based on deep learning is proposed.In this paper,considering that the noise and other interference factors in the data are sparse relative to the data itself,the traditional sparse representation algorithm and the deep learning algorithm are integrated to optimize the design of the total loss function.By adding a sparse term into the reconstruction loss,which is an error matrix,and it is sparse relative to the input data.Therefore,it can be achieved by The norm constrains the sparse term to train a subspace clustering network which is more robust to noise and singular points,so as to achieve better clustering effect.This algorithm not only improves the accuracy of clustering,but also improves the robustness of the algorithm to a certain extent.It has relatively stable performance when processing different data.2.A new clustering algorithm based on U-Net is proposed.Because the realistic scene data to be clustered generally contains various disturbance factors,and the most influential one is noise interference.At the same time,in recent years,the U-Net network has outstanding effect in data denoising.Therefore,this paper studies and draws lessons from it,considering that when clustering in subspace,if the images to be clustered are cleaner data,that is,the comparison of interference factors With fewer images,the clustering effect will be improved.Therefore,from the perspective of network design,a U-Net network branch is proposed to learn the noise and other interference factors in the image.Another branch,the deep auto-encoder network in this paper,is used to reconstruct the image and learn the selfrepresentation matrix.Then,the data are clustered on the basis of the self representation matrix.Through joint training,a pair of The performance of subspace clustering is greatly improved and the robustness is also significantly enhanced.The algorithm in this paper provides a good idea and reference for the combination of traditional algorithm and deep learning algorithm to solve the problem of subspace clustering.It also has great reference significance for the innovation of algorithm in this field.At the same time,the clustering effect achieved in this paper is also better in the same period.