Face Clustering Based On DBSCAN(Density-based Spatial Clustering Algorithm With Noise)

In this thesis following problem has been addressed: In a number of unlabeled face images data set,the images are clustered into the individual identities present in the data.This is a relevant problem in different applications from social media to law enforcement.Cameras surveillance are becoming increasingly popular in many organizations,from law enforcement,security,traffic control and residential.Police nowadays perform the investigation by searching for criminals or specifics individuals looking at videos or pictures taken from the cameras.Examinations done manually have become impossible therefore leading to some degree of automation.Density-based spatial clustering algorithm with noise(DBSCAN)algorithm will be focused in this thesis which is an efficient and effective way that shows good performance & it is considered as a good algorithm for face clustering providing better clustering accuracy than other well-known algorithms such as k-means,spectral clustering & Rank-Order when clustering faces.The core idea of the density-based clustering algorithm DBSCAN is that each object within a cluster must have a certain number of other objects inside its neighborhood.Compared with other clustering algorithms,DBSCAN has many attractive benefits,e.g.it can detect clusters with arbitrary shape and is robust to outliers,etc.therefore,DBSCAN has attracted a lot of research interest during the last decades with many extensions and applications.This research contributes with the comparison of several clustering algorithms which are described and applied to different data sets.The five clustering algorithms are k-means,threshold clustering,mean shift,DBSCAN,and Approximate Rank-Order.In the experiments,these cluster techniques are applied on subsets and the Labelled Faces within the Wild(LFW)data set.Also,a data set containing faces of people appearing in the database of ISIS is tested to evaluate the performance of these clustering algorithms.This research is intended to increase knowledge about the comparison of clustering algorithms for face clustering.The main finding is threshold clustering shows very good performance regarding the f-measure and number of false positives.Finally,DBSCAN has shown smarted and almost perfect clustering performance throughout experiments and is taken into account as an honest formula for face cluster.