P. Foggia, C. Sansone, and M. Vento (Eds.): ICIAP 2009, LNCS 5716, pp. 443–450, 2009. © Springer-Verlag Berlin Heidelberg 2009 Similarity Searches in Face Databases Annalisa Franco and Dario Maio DEIS – Università di Bologna, Viale Risorgimento, 2 – 40136 Bologna - Italy {annalisa.franco,dario.maio}@unibo.it Abstract. In this paper the problem of similarity searches in face databases is addressed. An approach based on relevance feedback is proposed to iteratively improve the query result. The approach is suitable both to supervised and unsu- pervised contexts. The efficacy of the learning procedures are confirmed by the results obtained on publicly available databases of faces. 1 Introduction Face is one of the most studied biometric characteristics and a huge literature exists about face recognition approaches [12]. Face recognition relies on the availability of a labeled database of faces, i.e. the identity of each face image is known, thus allowing the creation of a template representative of the user; recognition is performed by comparing the unknown input face image with the stored templates. In this work a different scenario is addressed where a similarity search is more appropriate than a direct identification. Several real applications fall in this category. For example in video-surveillance applications usually large quantities of images are gathered and stored to be examined in case of need (e.g. image sequences acquired at a bank en- trance could be analyzed after a robbery: an image of the robber could be compared with other images in the attempt of finding frames where the face is more clearly visible). Another important application in the law enforcement field is the typical mug-shot search problem where a witness has in mind the image of a subject and a support tool could help to leaf through a database of suspects. In this case a query image could not be available, so that typical face recognition approaches cannot be applied. Our framework was designed with the objective of providing a valid and flexible tool to perform similarity searches in face databases. The proposed system exploits relevance feedback techniques to gradually improve the search result. The search process is iterative and the information obtained at each iteration is used to improve the result in the subsequent search steps. The proposed system is very flexible being able to work in different modalities: supervised by the user (with or without a query image) and unsupervised. In the former, like in traditional feedback techniques [13], the system gathers information from the user judgment to improve the results, in the latter, which represents the main contribution of this work, an unsupervised iterative mechanism is exploited to improve the initial result without requiring the user inter- vention. The two modalities can either work independently or be combined to reduce the user effort needed to obtain a satisfactory result.