Rotationally and Illumination Invariant Descriptor Based On Intensity Order

Abstract

In this thesis, a novel method for local feature description where local features are grouped in normalized support regions with the intensity orders is proposed. Local features extracted using this kind of method are not only gives advantage of invariant to rotation and illumination changes, but also converts the image information into the descriptor. These features are calculated with different ways, one is based on gradient and other one is based on the intensity order. Local features calculated by the method of the gradient performs well in most of the cases such as blur, rotation and large illuminations and it overcome the problem of orientation estimation which is the major error source for false negatives in SIFT. In order to overcome mismatching problem, method of multiple support regions are introduced in the proposed method instead of using single support region which performs better than the single support region, even though single support region is better than SIFT. The idea of intensity order pooling is inherently rotational invariant without estimating a reference orientation. Experimental results show that the idea of intensity order pooling is efficient than the other descriptors, which are based on estimated reference orientation for rotational invariance.