ISIFT: Improving the performance of SIFT for mirror images

Feature detection is very important in computer vision field, handcrafted descriptors such as SIFT is widely used in image matching, image recognition and image retrieval. Although SIFT is scale and rotation invariant, but it is not reversal invariant. Therefore, we propose a new descriptor ISIFT (Improved SIFT) based on SIFT to improve the Performance. The descriptor combines the descriptor of both original image and reversed version. It has all advantages which SIFT have and it has reversal invariance at the same time. Also, RANSAC algorithm and distance ratio is used to remove mismatched key-points. ISIFT has much better performance in symmetric images, with this descriptor, an image can match both mirror-like image and itself-like image. At last, the experiment shows ISIFT have more correct key-point matches and higher accuracy in symmetric images than SIFT, MISIFT and ASIFT. Also, it performs as good as SIFT in scale transformation.