hrvatski

Neparametarska klasifikacija objekata na slici upotrebom konteksta

Segmentation and classification of image objects is one of the most important but also most complex problems in computer vision. Complexity of the problem is particularly boosted with images of natural landscape where there is huge variation of object appearance due to different perspectives, illumination and typically large number of partially visible objects. Human vision can handle natural images fairly, relying on strategies that exploit different sources of information, experience and learned rules about physical world. This is the reason that computer vision is trying to replicate same strategies used by human vision to improve systems at typical tasks of detection, location and classification of image objects. One of more important information sources is image context composed of different sources within or outside the image. Heterogeneity of contextual information sources is the cause of many approaches for image object classification that integrate context and object appearance. In this work we propose new model for natural image object classification using context at the level of image segments. Context modeling is largely independent of classification based on appearance and proposed model enables simple upgrade of existing systems with information from global and/or local context. Context modeling is based on nonparametric use of appearance classification results which is novel approach compared to previous systems that model context on limited number of rules expressed with fixed set of parameters. Model implementation resulted in system that in our simulations showed stable improvement of object classification based on appearance.

klasifikacija objekata; prirodna slika; kontekst slike

nije evidentirano