Merging and Generalizing Eigenspace for Partially Occluded and Destroyed Object Recognition

An eigenspace generalization technique for overcoming partial occlusion and destruction of an object is proposed in this paper. Eigenspaces of various partially occluded and destroyed shapes of a particular object are merged with the corresponding eigenspaces of good-shapes in order to reduce the effect of data-loss due to occlusion and/or destruction. If an eigenspace is developed with various poses of objects, similar poses store together with respect to their appearances. We take this advantage and generalize this window (namely eigenwindow) by merging and averaging their feature-points in each and every window. This generalized or mean eigenwindow is further used for recognizing an unfamiliar pose, including partially occluded and/or destroyed shapes, and the object type itself. We have applied the proposed approach to various data-loss environments and the method has successfully performed the recognition of an object with up to 20% of occlusion and/or destruction. An extensive experiment is also performed and recommended for overcoming the background effect.