Abstract:  An image stitch method in video scene need to perform well in real time, scale and rotation invariance. Aiming at the issue that the ORB algorithm does not have scale invariance and SIFT algorithm with high time complexity does not respect the natural boundaries of objects, we put forward an image registration algorithm based on the nonlinear scale space and improved the ORB and a mean coordinate blend method based on YUV brightness template. Firstly, the nonlinear scale space is used to smooth noise and retain object boundaries, to set scale parameters based on image entropy and to set proper distance between feature points, so that the ORB detectors are stable. Then, maxima is searched for in scale and spatial location and descriptors are compute. Finally, the feature points are matched with hamming distance and RANSAC algorithm. The experimental results show that the improved algorithm can improve robustness, shorten time of registration and improve greatly the matching accuracy. What’s more, the mean coordinate method with brightness template enhances blend.

Key words: image registration, blend, nonlinear scale space, image entropy, brightness template