Object Recognition using Full Pixel Matching

We consider the optimization problem of object recognition for real world images. Although several approaches have been proposed, this paper aims to improve the recognition rate with our novel method. In this paper, a full pixel matching based object recognition method in which no advance segmentation procedure is occurred during matching procedure is proposed. Our method compares the similarity of two images with pixels rather than the previous works with regions, shapes, et. al, so that the recognition rate can be improved. Furthermore, in order to implement our method, we present and analyze two algorithms which are Decision Space based Algorithm (DSA) and Direction Pattern based Algorithm (DPA). In the experiment, the performance of recognition of our two algorithms is evaluated on caltech 101 dataset. We compare our algorithms to several conventional works, such that our method improve the performance of object recognition in recognition rate, robustness of recognition among the variation of appearance and deformation in the images, and segmentation free.