Radar recognition is an important task in many flelds. The diversity of image sources requires the recognition be invariant to many kinds of image transformations. SIFT is proved to be the best local descriptor in achieving invariance. According to the fraction of keypoints that are matched to the reference image, two thresholds are set to provide more accurate recognition. A soft decision with a confldence degree is provided to satisfy various demands for difierent accuracy. Experiments show that SIFT achieves good invariance to scale, brightness, rotation and slight a‐ne transformations.