Research On Image Local Feature Matching Algorithm And Its Applications

Image local invariant feature, which plays an important role in pattern recognition, image processing and computer vision tasks. Since the image local features are invariant to a lot of image transformations like rotation, viewpoint, scale, illumination, blur, compression and so on, they have been intense researched and demonstrated to be successful in a wide range of applications, such as stereo matching, image registration, image stitching, object recognition, image retrieval and target tracking.Recently, with the development of computer vision technology, more and more applica-tions require image local invariant features. Meanwhile, different applications demand for var-ious image invariant features. Therefore, there are still many problems to be improved and resolved in the research on local invariant features. In this dissertation, in order to improve the performance of classification and recognition based on invariant features, some theories about invariant features were analysed thoroughly and some existing methods of extracting and matching invariant features have been studied. Then several novel algorithms based on the state-of-the-art ones with better performance in stereo matching, image registration, remote sensing image processing and target tracking have been proposed.First, a rotation-invariance improved SURF algorithm was proposed. To solve the classical SURF’s poor performance on the rotation invariance, this dissertation proposes a new matching algorithm combining the SURF feature point and DAISY descriptor. The proposed method utilize the fast Hessian matrix approximation to detect feature points. In addition, we propose a main orientation distribution method which is more suitable for DAISY descriptor, so that a new descriptor can be obtained via rotating by the main orientation. Our algorithm effectively improves the matching ability of the classical SURF algorithm on the rotation invariance but only employs a little more computational burden. The experimental results demonstrate that our algorithm is more robust than classical methods when the image blur, illumination, JPEG compression ratio or the viewpoint changes.Meanwhile, an extended SURF descriptor was proposed and adopted effectively in the application of remote sensing images registration. To solve the classical methods’problems of long executing time or low accuracy, this dissertation proposes an extended SURF descriptor. On the method using SURF, the proposed method uses local normalization information and second-order gradient values of neighbourhood regions to build a new one. Not only does this method perform as fast as SURF algorithm, but it also fully employs the image gray-scale information and details. Considering the executing time and rate, experimental results presented in this dissertation show that the proposed method is more robust than classical methods.Furthermore, due to the local information ambiguities of the images containing repetitive patterns, false matches can be easily produced by the local feature based image matching al-gorithms. The matching algorithms combining with the global feature still depend on the main orientation which is obtained by calculating the local information. Therefore, these algorithms also usually lead to mismatching for the images with repetitive patterns. Thus, it is meaningful to cope with the challenging matching task since such repetitive patterns widely exist in the real world images of artificial objects or scenes. To solve this problem, a novel image matching algorithm based on pair-wise feature points is proposed in this dissertation. The direction vec-tor between the pair-wise points is utilized to be the main orientation, which provides the right direction for both the local and global feature description. In addition, local DAISY descriptor and the improved global context descriptor are used in the proposed algorithm to improve the matching ability. We evaluate the proposed method on both the simulative and real images a-gainst several state-of-the-art algorithms. Experiments on images of both simulative and real as well as comparisons with the state-of-the-art methods have demonstrated the effectiveness and robustness of the proposed method. Moreover, the proposed algorithm is an effective approach to solve the repetitive patterns images matching problem.Finally, visual target tracking is a primary task in many computer vision applications and has been widely studied in recent years. Among all the tracking methods, the mean shift al-gorithm has attracted extraordinary interest and been well developed in the past decade due to its excellent performance. However, it is still challenging for the color histogram based algo-rithms to deal with the complex target tracking. Therefore, the algorithms based on other dis-tinguishing features are highly required. In this dissertation, we propose a novel target tracking algorithm based on mean shift theory, in which a new type of image feature is introduced and utilized to find the corresponding region between the neighbour frames. The target histogram is created by clustering the features obtained in the extraction strategy. Then, the mean shift process is adopted to calculate the target location iteratively. Experimental results demonstrate that the proposed algorithm can deal with the challenging tracking situations such as:partial oc-clusion, illumination change, scale variations, object rotation and complex background clutter. Meanwhile, it outperforms several state-of-the-art methods.