In order to improve the robustness and real time performance of SIFT based image registration algorithms, a new
descriptor is proposed. We compute the new descriptor for a log-polar location grid with 3 bins in radial direction (the
radius set to 3, 6 and 8) and 12 in angular direction, which results 36 location bins. Firstly, the 3×12 log-polar location
grid is rotated to align its dominant orientation to a canonical direction in a different way with SIFT with less
computational complexity. The keypoint dominant orientation and its orthogonal orientation is defined as the x and y
directions of the descriptor’s local coordinate system. After that a 3×3 grid is rotated to align the dominant
orientation ,and then the rotated 3×3 grid is translated to each pixel within the local 3×12 log-polar location grid.
Therefore the local 3×12 log-polar location grid is rotated to align the dominant orientation and the rotation invariance is
achieved. Secondly, compute the gradients in x and y directions using the rotated 3×3 grid for each pixel within the
keypoint local neighborhood, so each pixel constitutes a gradient vector with 2 dimensions. Thirdly, use the distance
from each pixel to the corresponding keypoint and a Gaussian function to assign a weight to the gradients, and then
compute the gradients location histogram for the 36 location bins in the 3×12 log-polar grid. Finally, a descriptor with 72
dimensions is constituted and the descriptor is normalized to achieve illumination invariance. The experimental results
show that the computational complexity of the new descriptor is reduced about 30% compared with the standard SIFT
descriptor while the performance is favorably compared to the standard SIFT descriptor .