Object Recognition with Discriminative Elastic Grids

The following paper presents a concept of Discriminative Elastic Grids (DEG) – an object recognition approach derived from the Deformable Template paradigm. DEG is an array of appropriately preset nodes, which interact with each other and with an image. Node information is derived to ensure distinctive grid deformations for objects of different categories. Grid deformations are summarized using descriptors, determined using supervised-PCA analysis of matched-grid node displacements. Speech-reading hasbeen selected as an application for experimental verification of the proposed concept. It has been shown that DEG-based representation of phone images provides high recognition rates for the considered, difficult image processing task.