Affine invariant matching of noisy objects.

Abstract

In computer vision many techniques have been developed for object recognition. The affine invariant matching algorithm proposed by Hummel and Wolfson (1988) is a new and interesting method. Under affine invariant transformation, objects with translation, rotation, scale changes, and, or even partial occlusion will have the same or similar coefficients. However, some serious problems exist in the original algorithm. This thesis begins with the discussion of the affine transformation. The shortcomings that can occur in this method such as the basis instability, the collision of hash table, and the noise sensitivity will be discussed. Among them the noise sensitivity is a serious problem. This can always cause the recognition procedure to fail. In this thesis an improved affine invariant matching algorithm was developed to overcome the noise problem and other disadvantages of the original algorithm. The area test criteria were adopted to avoid the numerical instability problem. The modified hashing structure using a special hash function was implemented to achieve faster accessing and voting. In the recognition procedure, the partial voting technique with the consideration of false peaks from the voting array highly enhanced the noise tolerance of the algorithm. Finally, the results obtained from the improved algorithm clearly showed better performance than those of the original algorithm.