Naučno-tehničeskij Vestnik Informacionnyh Tehnologij, Mehaniki i Optiki (Jul 2015)
PEOPLE RETRIEVAL BY MEANS OF COMPOSITE PICTURES: METHODS, SYSTEMS AND PRACTICAL DECISIONS
Abstract
We discuss the problem of people retrieval by means of composite pictures and methods of its practical realization. Earlier on, the problem was posed in the previous paper by the authors, and this paper deals with its further development. The starting premise here is that for the successful search of people by their sketches it is necessary to transform these sketches into sets of populations of sketches imitating evidence of «group of witnesses» and evidence with incomplete information in verbal portraits. Variants of structures for benchmark «photo-sketch» databases are presented, intended for modeling and practical realization of original photos retrieval by sketches, which new component is a population of sketches. Problems of preprocessing for initial sketches and original photos and its influence on the result of their comparison are discussed. Simple sketch recognition systems (Simple FaRetSys) and a problem of original photos retrieval by the sketches are considered. Shortcomings of such systems are shown and new decisions on extending and development of simple systems (Extended FaRetSys) are presented. Experiments on searching of original photos by sketches in the CUFS database of sketches and similar experiments on widely known FERET and CUFSF facial databases are presented. Three frameworks are offered for retrieval performance improvement. In the first one, original sketches are transformed into populations, and then in these populations the sketch similar to the given sketch (Forensic Sketch) is already defined. The class of the sketch found in a population «by definition» unambiguously corresponds to a class of the original photo. In the second framework, the Forensic Sketch is transformed to a population of sketches, and all original sketches in a benchmarking database are compared to sketches from populations of the Forensic Sketch. The class of matches is determined in the same manner as in the first framework. The third framework includes generation of a population of sketches, both from all original sketches, and from all Forensic Sketches. The further line of research is obvious: retrieval by matching between sketches of these two populations.
Keywords
