Image registration algorithm based on gradient descent

Image registration is a key technique in image analysis. This paper begins with an analysis of the shortcomings of the Lucas-Kanade algorithm and the existing improved algorithms. Aiming at the shortcomings of these algorithms, that is, huge computational cost, an image registration algorithm based on gradient descent is presented. Firstly, the algorithm redefines the objective function by switching the role of the image and the template. Then, the Gauss-Newton gradient descent algorithm is used to get the increments of the parameter. Since there is nothing in the Hessian matrix that depends on the parameter, it is constant in every iteration and can be pre-computed. Finally, the parameter is iteratively solved until it satisfied the test for convergence. Experiments with several standard sequences and using the set of affine warp which is adapted to any combinations of the template's rotation, zooming and translation indicate our new algorithm's capability. With the same description of the template it achieves the same precision as the Lucas-Kanade algorithm. The comparison of computational cost between our algorithm and the Lucas-Kanade algorithm demonstrates the improvement of our algorithm.