Distinctive image features from illumination and scale invariant keypoints

This paper proposes a novel local feature descriptor of the image, which is named iSIFT (illumination and Scale Invariant Feature Transform), based on SIFT (Scale Invariant Feature Transform) improved by LBP (Local Binary Pattern), in order to combine the robustness advantages of LBP descriptor for illumination change and that of SIFT for scaling. It addresses the following problems: (1) SIFT algorithm is poor in describing the local feature extraction from an image when lighting condition changes; (2) SIFT algorithm cannot accurately extract the feature points or can only extract only few of them from the blurred image and the image of an object with smooth edges. Each of the scale-space representation, namely, L(x, y, k sigma), in Gaussian pyramid of the image I(x, y) on SIFT descriptor is calculated by using LBP in order to obtain the corresponding LBP image, which is denoted by LBP(L(x, y, k sigma)). The obtained LBP(L(x, y, k sigma)) replaces the original corresponding scale-space representation L(x, y, k sigma) to construct the LBP-Gaussian pyramid, and the difference between each two neighboring LBP(L(x, y, k sigma)) in LBP-Gaussian pyramid is used to replace the original DoG pyramid in SIFT descriptor to detect extreme points. The results of the experiments suggest that iSIFT descriptor improves the precision of image feature matching and the robustness under changed lighting conditions compared with that of SIFT algorithm, and iSIFT descriptor can extract more feature points from the blurred image and the image with smooth edges as well as having stronger robustness for lighting, rotation and scaling.