LAUSR

Dense matching is the basis for many advanced image processing algorithms such as 3D reconstruction, super-resolution reconstruction, and image fusion. However, it has several limitations in speed and accuracy; the main aspects affecting the practicality of dense matching include the space–time efficiency, robustness, and matching accuracy. To overcome these limitations, a robust dense matching algorithm based on sparse matching is proposed. First, a sparse matching algorithm is used to obtain a sparse point set. Then, based on the rotation angle, the scaling ratio, etc., the neighbourhood centre, the neighbourhood size, and the extraction order of the descriptors are determined, and the camera pose is calculated to evaluate the confidence of the matching points in the sparse matching point set, and the outer points with lower confidence are excluded. The HOG descriptor in the neighbourhood is extracted and convolved to obtain a fractional matrix, and the gradient variance of each pixel in the neighbourhood is added to the fractional matrix with certain weight. Finally, the new matching points are determined by the obtained fractional matrix and the given number of directions to achieve the densification. The experimental results showed that the proposed method can achieve fast and high-quality dense matching for images with repeated texture, rotation, scaling, and affine transformation and had high space–time efficiency and accuracy.