Object Recognition by Modified Scale Invariant Feature Transform
Gul-e-Saman, S.A.M. Gilani · 2008
This paper presents a methodology for object recognition. It relies on the extraction of distinctive invariant image features that can be used to find the correspondence between different views of an object or a scene. These features are invariant to image rotation and scaling, they have substantial robustness to changes in viewpoint and illumination and addition of noise. Mikolajczyk [1] have evaluated the SIFT [2] algorithm along with other approaches and have identified it as the most resistant to image distortions. This paper improves on the SIFT algorithm by modifying its descriptor and the keypoint localization steps. The proposed technique uses the salient aspects of image gradient in keypoints neighbourhood. Moreover, instead of smoothed weighted histograms of SIFT, kernel principal component analysis (KPCA) is applied in order to normalize the image patch. Comparative results show that KPCA based descriptors are more distinctive, robust to distortions and compact. The evaluation of the technique is performed using recall precision [3].