A method of generating Key.Net sub-pixel key-points by local gradient fitting

In this paper, we propose a new approach for sub-pixel registration based on Key.Net combined with a local gradient fitting method. Considering Key.Net of pixel-level accuracy cannot meet the need of high-precision image matching, we innovatively calculate the gradient of the response values of key-point and its neighbors in the response map obtained from the Key.Net, which effectively avoids errors caused by factors such as illumination and shadow for gray value calculation by traditional gradient methods. We first apply the Key.Net detector to extract multi-scale key-points from input images, then build response a map using response values of all pixels on each scale plane and calculate the sub-pixel offset of a pixel-level key-point based on the local gradient relations of the key-point and its neighbors in the response map. Finally, we get the sub-pixel coordinates of the key-point. The experimental results show that the proposed method successfully matched some low-quality images and detected large numbers of correct matching points. The localization accuracy of key-points was improved to 0.1 pixels, while the excellent feature detection ability of Key.Net was retained. When testing using the HPatches sequences dataset, our method achieved the highest matching score for view-point, which is 15.5% higher than SIFT, and 0.6% higher than Key. Net-HardNet after increasing the number of scale layers.