Research Of Image Matching Technology Based On Local Feature Detection

Traditional image matching algorithms can not be used to deal with complex image data.Practice shows that the result of matching will be better if it is combined with the image features.The local features of image, which would not be affected by the environment easily,have been applied to the matching algorithms increasingly. How to improve the accuracy and efficiency of image matching algorithms based on local feature has been one of hot topic of image matching. This thesis studies image matching technology based on local feature. The main work is as follow:(1) We present a local feature matching algorithm based on Contourlet transform to improve the accuracy. In this method, Canny edge detection algorithm is used to remove the edge extremal point of DoG space. By this way, we remove the unstable edge extremal points and ensure the uniqueness.There comes Contourlet transformation of these key points which can decrease the dimension of feature descriptor and introduce global texture information to the SIFT local features. This algorithm can track similar areas accurately. Finally, the combination of cityblocks distance and chessboard distance is employed to measure the feature descriptors. Simulation results indicate that the accuracy of this algorithm is higher than SIFT.(2) A method called two-column histogram hashing is presented to reduce the computational complexity of feature points extraction. Then combine it with ORB algorithm to proposed a image matching method using a novel two-step searching strategy(coarse-to-fine). The proposed method is combined with two-column histogram hashing and ORB. Firstly, coarse matching is conducted with a novel two-column histogram hashing. Then, the approximate location where target object may appear is analyzed. Secondly, in the refining stage, key points are detected and described in location above using ORB. The result of coarse matching can narrow the scope of feature extraction in refining stage and save time of image matching. Experimental results imply that the proposed approach outperforms SIFT, SURF and ORB algorithms in speed of image matching.(3) Traditional local feature matching algorithms ignore the color information of images and have high error rate when a lot of similar regions are matched. This thesis proposes a improved method, which combining the color invariance with shape context of images. The proposed approach takes color invariance as the input images and builds descriptors with the improved GLOH method. Then, to decrease the match error rate, the global shape context descriptors isgenerated by detecting the principal curvatures of the feature points. Finally, it ingretates two descriptors above. The evaluation represents that this method provides higher accuracy than other three related methods of previous research in many cases, such as similar local regions, illumination change and image blur.