Corresponding feature extraction is a key stage in infrared-visible image registration, fusion, and change detection, etc.. Considering the difficulties in correct extraction of related features between infrared and visible images of the same scene, an affine invariant method is proposed based on Maximally Stable Extremal Regions (MSER) algorithm. The approach includes three steps: (1) to extract the maximally stable extremal regions in infrared and visible images; (2) to fit the feature regions to ellipses; and (3) to regularize the elliptical regions to eliminate the deformation disturbance. The final output is coherent features which are convenient for description and matching. Experimental results show appealing effectiveness of the proposed method in corresponding feature extraction between infrared and visible images.