ATTRIBUTED RELATIONAL SIFT-BASED REGIONS GRAPH (ARSRG):DESCRIPTION, MATCHING AND APPLICATIONS

Finding correspondences between images is a crucial activity in many overlapping fields of research, such as Image Retrieval and Pattern Recognition. Many existing techniques address this problem using local invariant image features, instead of color, shape and texture, that to some degree loose the large scale structure of the image. In this thesis, in order to account for spatial relations among the local invariant features and to improve the image representation, first a graph data structure is introduced, where local features are represented by nodes and spatial relations by edges; second an algorithm able to find matches between local invariant features, organized in graph structures, is built; third a mapping procedure from graph to vector space is proposed, in order to speed up the classification process. Effectiveness of the proposed framework is demonstrated through applications in image-based localization and art painting. The literature shows many approximate algorithms to solve these problems, so a comparison with the state of the art is performed in each step of the process. By using both local and spatial information, the proposed framework outperforms its competitors for the image correspondence problems.