Cluster Analysis And Its Applied Research On Camera Source Identification

As an unsupervised learning method,cluster analysis is one of the important research methods in the field of machine learning.It has been successfully applied to many fields,such as finance,business,social network and bioinformatics.There are a lot of effective clustering algorithms currently,spectral clustering based on graph theory has received wide attention due to the advantage of being able to cluster data with arbitrary shape and easy implementation.However,spectral clustering is inapplicable to large-scale datasets due to its high computational complexity and large space overhead.Sensor pattern noise is seen as an effective and stable camera fingerprint,it can be used for camera source identification and image tamper detection.This paper studies the scalability of spectral clustering algorithm and its application on camera source identification for large-scale image datasets.The main contributions are listed as follows:(1)The existing semi-supervised spectral clustering algorithm can neither effectively deal with large-scale datasets nor make full use of the limited constraint information,hence a semi-supervised spectral clustering combining sparse representation and constraint propagation is proposed.The data points in the constraint set are seen as landmarks and used to construct the sparse representation matrix,the graph similarity matrix is obtained approximately,so as to improve the constrained spectral clustering model and improve the scalability of clustering algorithm.The connected regions can be generated based on the similarity matrix of landmarks,the neighbor nodes are dynamically adjusted in each connected region,clustering accuracy is further improved by using constraint propagation.Experimental results show that the proposed algorithm can significantly reduce computational complexity without obvious accuracy decline and can be scalable to large-scale datasets.(2)Spectral clustering algorithm suffers from high computational complexity for large-scale datasets,even some works proposes to replace the eigen-decomposition with stacked autoencoders,then ran k-means clustering on embedding representation,which improved efficiency but further increased the memory consumption by using the similarity matrix of all data points.We propose to select landmarks and use the similarity of landmarks with other data points as the input of deep autoencoders instead of similarity matrix of the whole datasets.The clustering loss based on KL divergence is used to update the parameters of autoencoders and cluster centers simultaneously,and reconstruction loss is also included to prevent the distortion of embedding space.Experimental results demonstrate that the proposed algorithm can achieve better clustering performance compared with traditional improved spectral clustering and deep clustering algorithms.(3)Aiming at the problem that most camera source identification methods based on clustering need to compare the high dimensional camera fingerprint and matrix decomposition,a fast image clustering algorithm based on camera fingerprint is proposed.The camera fingerprints are appropriately compressed by using random projection to reduce the computational complexity of fingerprint matching.While the improved spectral clustering methods are applied to camera source identification on large-scale image datasets,we adjust sparse representation matrix according to the normalized correlation coefficient between camera fingerprints.Meanwhile,pairwise constraint information between some images is established to guide clustering process.Experimental results show that the proposed methods can be applied to large-scale image datasets,the accuracy of camera linking is guaranteed as well as improving the efficiency.