Topological Gabor descriptors: exploring a filter bank structure for image feature matching

In this paper, we propose a novel feature descriptor based on Gabor filters, called Topological Gabor Descriptor. We build a filter bank in such a way that the descriptors are invariant to rotation and scale changes. The filter bank topology enables a simple matching scheme based on the circular shift of descriptors. We evaluate the effectiveness of our approach to feature description in an object/scene recognition setting. The descriptors were evaluated with synthetic and real images. The performance of the descriptors was measured by computing the average matching rate. Our experiments with synthetic data show a robust invariance property for a high degree of rotation and scale variations. Our experimental results shows a 93.50% matching rate for synthetic images subjected to rotation. The matching rate for a scale variation of up to two times the original scale is 81.11%. The methods discussed in this paper were also tested on three different datasets of real images of buildings where we obtained an average matching rate of 41.33%.