This thesis describes a complete framework for organising digital photographs in an unsupervised manner, based on the appearance of people captured in the photographs. Organising a collection of photographs manually, especially providing the identities of people captured in photographs, is a time consuming task. Unsupervised grouping of images containing similar persons makes annotating names easier (as a group of images can be named at once) and enables quick search based on query by example.
The full process of unsupervised clustering is discussed in this thesis. Methods for locating facial components are discussed and a technique based on colour
image segmentation is proposed and tested. Additionally a method based on the Principal Component Analysis template is tested, too. These provide eye locations required for acquiring a normalised facial image. This image is then preprocessed by a histogram equalisation and feathering, and the features of MPEG-7 face recognition descriptor are extracted. A distance measure proposed in the MPEG-7 standard is used as a similarity measure.
Three approaches to grouping that use only face recognition features for clustering are analysed. These are modified k-means, single-link and a method based on a nearest neighbour classifier. The nearest neighbour-based technique is chosen for further experiments with fusing information from several sources. These sources are context-based such as events (party, trip, holidays), the ownership of photographs, and content-based such as information about the colour and texture of the bodies of humans appearing in photographs. Two techniques are proposed for fusing event and ownership (user) information with the face recognition features: a Transferable Belief Model (TBM) and three level clustering. The three level clustering is carried out at “event” level, “user” level and “collection” level. The latter technique proves to be most efficient.
For combining body information with the face recognition features, three probabilistic fusion methods are tested. These are the average sum, the generalised product and the maximum rule. Combinations are tested within events and within user collections. This work concludes with a brief discussion on extraction of key images for a representation of each cluster.