In Conclusion
It was shown that the method of classification considered here can perform better than the moments method on certain kind of noise (non-uniform distortions and dilation), in exchange for sacrificing performance on other types of noise.
It was found that the recognition algorithm performance depends on an evolution parameter as well as “complexity” of the original image. Another experiment to be done is to set the evolution parameter to depend on the image “complexity”, or compactness, or even to halt evolution of the original by some threshold on distance (in tangent space representation) between the polygons of two consecutive steps of evolution.
Another way to improve the performance is to use a similarity measure proposed by Latecki et al [5]. That measure introduced is related to the measure used in this work. The main difference is that the measure L2 is applied to the pairs of polygonal shapes, which makes the approach more robust with respect to noise and non-uniform distortions. Also, the performance of the algorithm on different kinds of noise and distortions should be examined in greater detail.