Scient<strong>ific</strong> Bulletin – Economic Sciences No. 7 (13)Multimedia system for interactive visual recognition of computer-generatedobjectsUniversity Lecturer PhD Florentina ENESCUUniversity Professor PhD Mariana JURIANUniversity of Pitestienescu_flor@yahoo.comAbstract:In this paper we present a project <strong>de</strong>veloped on the Windows platform, integrated into thelocal network(LAN) and implemented on a PDA (Personal Digital Assistant) and aportable computer configured as a mobile customer workstation in a wireless network (NFO or both Wimax), both with installed web cams .Keywords: Digitized multimedia, feature <strong>de</strong>tection, relevance judgements, multimediadatabase, query processing, content – based retrieval, data mining, image in<strong>de</strong>xing andretrieval, metadata, similarity, image clustering, electronic publishing.1. IntroductionThe difference between this system and the classical frequent-used systems is theinteraction with the operator based on a real mo<strong>de</strong>l. The result of an application is presented: thereshaping and retrieval of the images from a database, the advantages of the architecture of somerecognition systems are commented. The systems of visual automatic recognition rarely obtain100% correct assortment on family of interest objects. Most of them permit the interaction withuser in the beginning, localized and <strong>de</strong>clared, not acceptable in scient<strong>ific</strong> articles. Delivering theways of interaction with the process unit, this is efficient from the viewpoint of the timeconsumed. Leaving the operator completely check of the process of analysis is much more thanif we seek the answers to general questions regar<strong>din</strong>g the computer. The mobile recognitionsystems offer, obviously, the advantages in the recognition of objects far from his position (likeflowers, other images). The first advantage of the method presented consists in the fact that itpermits to achieve image <strong>de</strong>finition of a more complex object. Then, the analysis may be basedon some images and the class that obtains the greatest probability is accepted. Methods thesophisticated processing can be <strong>de</strong>veloped for the <strong>de</strong>tection of relevant information from images,features, or the level of class<strong>ific</strong>ation.2. The method of thingAs in all analysers, a set of images with reference labels, split in the many classes, arestored in the database. Automatically, the algorithms segment each unknown image, build avisible mo<strong>de</strong>l, and draw upon the images of unknown object, a set of distinct pre-programmedfeatures that can be compared with formal features, colour or texture extracted from the thatimages. Applicants are then classified automatically accor<strong>din</strong>g to the similitu<strong>de</strong> features, withthat the unknown image.If one dyne the candidates display is similar to the unknown image, the operator selects themsimply by clicking on it, classifying in this unknown class. By that, the operator can adjust thevisible mo<strong>de</strong>l. The visible mo<strong>de</strong>l gui<strong>de</strong>s the system in the extraction of the features. Every timethe visible mo<strong>de</strong>l is built, features are extracted and all candidates are automatically registered.78
Multimedia system for interactive visual recognition of computer-generated objectsSometimes the correct candidate isn’t displayed until the adjustment of the visible mo<strong>de</strong>l isdone. In this case, the operator can explore the less sign<strong>ific</strong>ant candidates by clicking on the Nextbutton. If we have multiple images available as reference for the class, the operator also canexamine them. The methodology is represented in fig. 1.MODELUnknown objectEXTRACTINGFEATURESMod<strong>ific</strong>ationSIMILITUDEDataBase with imagesOK ?CLASSIFYYesFig. 1 Flowchat of the system. The human shares the by-path writed with redWhen a new image of an unknown object is taken, the algorithmic part of the system classifiesthe candidates through feature comparison from the new images, using all other images as areference. The candidates of the new set are displayed (fig. 2: a visible mo<strong>de</strong>l built by theautomaton from the unknown object).79