In order to solve the registration problem of retinal images with changes in the angle of view of the fundus camera, a local feature registration method with affine invariance is proposed. First, an affine-invariant anisotropic image structure is obtained by computing the minimal boundary rectangle of the connected domain corresponding to the vascular intersection point. Then, the anisotropic image structure is transformed into an isotropic image structure with different angular factors by rotating, sampling, and compression. Finally, using the feature invariant feature transform algorithm to extract feature points on the isotropic image structure, and the feature point coordinates are matched. The experimental results in a private sample set containing 83 groups of retinal images show that the root mean square error of the method is 1.247±0.251 pixels, which can solve the problem of perspective and scale change well.