Research on Improved AffNet Image Feature Matching Algorithm

Abstract

Aiming at the disadvantages of the commonly used local feature matching algorithms, which are less stable and rely on manual production of descriptors, this paper proposes a local feature matching algorithm based on deep learning. In order to better protect the image edge and detail information, a nonlinear filtering algorithm is used to construct a nonlinear scale space, which can effectively increase the stability of feature point detection and extraction compared with a Gaussian scale space. In order to make deep learning with better spatial transformation ability, STN spatial transformation convolutional network is added to the AffNet model, which can effectively prevent the information loss caused by spatial transformation. The proposed model is trained in the HPatch dataset, and the ability of the proposed algorithm in anti-affine transformation, illumination transformation, scale transformation, etc. is judged with the Oxford dataset. The proposed algorithm can be widely used in image stitching and other fields.