Object oriented image analysis based on multi-agent recognition system

In this paper, the capabilities of multi-agent systems are used in order to solve object recognition difficulties in complex urban areas based on the characteristics of WorldView-2 satellite imagery and digital surface model (DSM). The proposed methodology has three main steps: pre-processing of dataset, object based image analysis and multi-agent object recognition. Classified regions obtained from object based image analysis are used as input datasets in the proposed multi-agent system in order to modify/improve results. In the first operational level of the proposed multi-agent system, various kinds of object recognition agents modify initial classified regions based on their spectral, textural and 3D structural knowledge. Then, in the second operational level, 2D structural knowledge and contextual relations are used by agents for reasoning and modification. Evaluation of the capabilities of the proposed object recognition methodology is performed on WorldView-2 imagery over Rio de Janeiro (Brazil) which has been collected in January 2010. According to the obtained results of the object based image analysis process, contextual relations and structural descriptors have high potential to modify general difficulties of object recognition. Using knowledge based reasoning and cooperative capabilities of agents in the proposed multi-agent system in this paper, most of the remaining difficulties are decreased and the accuracy of object based image analysis results is improved for about three percent.

► Our object recognition method is composed of two main steps.
► First, object oriented scene analysis and second, multi-agent system.
► Contextual relations have high potential to improve recognition of urban objects.
► Structural descriptors modify general difficulties of recognizing objects.
► Capabilities of agents in multi-agent system improve overall accuracy of results.