The Study On Local Invariant Features Of Images

Feature matching among images is the foundation of most applications in machine vision. The images may have not only linear geometric deformation, such as scale, rotation and so on, but also illumination condition changes or viewpoint changing as well as located in cluster background or image occlusions. The method of local invariant feature overcomes features expression from image transformations, thus becomes a very useful tool under the visual circumstances changing. At the same time, due to emphasis on some kind of features in local area, it has the characteristics of low redundancy in the efficiency and speed. This method has become a very active research direction, and there are a lot of research results applied in machine vision, for example object recognition, image matching, image retrieval, robot navigation, scene panorama and 3D reconstruction. There are two problems needed to be solved in this method: the first one is location of the characteristics area, called detector algorithm; the second one is information description in the area, called descriptor algorithm. Choosing image matching in robot navigation system as applications background, this thesis concerns with the efficiency of the detector algorithm as well as the robustness in image transformations, and high efficiency and invariance to viewpoint changing of descriptor, then proposes a novel fast detector algorithm and descriptor algorithms, which can all be used in real-time applications. On the other side, this thesis also discusses inspiration the recognition mechanism of biological visual cortex. A new feed forward hierarchy image classification model is proposedFirst, for detecting the local feature points in a accurate, robust and efficient way, this thesis surveys the research status of fast detector algorithm and analysis the several classical fast detector algorithm, such as DoG,FAST as well as Fast-Hessian. According the performance analysis of existed detectors, we know that the detector algorithm based on Hessian matrix is the most time-saving one, which is better than other algorithms on stability and accuracy of features. A novel detector algorithm based Hessian matrix was proposed in this thesis, using mean filter to estimate Gaussian second order partial derivative, avoiding scale space iterative computations. In addition, using trace and determinant of Hessian matrix as criterion simultaneously to extract interest points can get rotation invariance, and can avoid the extracting feature points unstably from noise or texture image.Second, a local feature description method based on image pixel distribution is proposed in this dissertation. The performance of descriptor based on distribution is better than others. From the pixel intensity distribution, through the relatively simple intensity difference compared to build local description, and use the lower-dimensions descriptor to shorten the feature matching time, and establish a new descriptor algorithm based on pixel distribution, which is suitable for scale change, plane rotation and linear illumination change to a certain limit. After normalizing pixel intensity of local neighbor areas surrounding interest points, descriptors could be computed based on the intensity distribution property of local neighbor. During the grayscale differences calculation of based on normalized segmentation point, removing pixel intensity differences caused by the illumination changing, that makes features description satisfy invariant under illumination transformations.In addition, image matching in mobile robot navigation path applications, often face various deformations except objects rotating or scale zooming. This could be estimated by affine transformation. The existing affine invariant mostly depends on detector algorithm extraction local features of affine invariant, which is not suitable for many real-time applications. The reason is extracting affine invariant features from the whole image will causes large computing costs. Studying the description algorithm for affine invariant of local features can decrease greatly the whole time for local feature extraction. As a powerful tool in image processing, moments'research history and the fundamental theory are reviewed in this thesis. Invariant moment as well as the affine invariant moments are introduce later, and then a novel descriptor based on Flusser&Suk moments is proposed——the affine invariant moment based on local gradient. In order to ensure the invariant of description features during illumination transformation, when extracting the central moment in local area, normalized local gradient algorithm is adopted without depending on original pixel values, which could reduce degree of dependence to the original pixels intensity and improve robustness of the description in the illumination transformation.Finally, this thesis applies local features in image classification and recognition. The original source of local features is consistent with identification mechanism of biological vision. The vision receptive proposed by Wiesel and Hubel is a natural processor for local features. So image classification based on local features is in accordance with the features-dealing mechanism of biological visual identification. On the other hand, image classification system based on biological visual cortex of hierarchical model is the hotspot in image classification researching. A new feed forward hierarchy image classification model is proposed in the article, applying the latest neurophysiologic findings about visual cortex recognition and local feature method. In order to enhance the stability of image features in multi image transformations, the primary features are processed in sparseness mechanism. At the same time, this method overcomes the disadvantages existing in formal models, such as poor robustness in rotation as well as illumination changing, a large amount of primary features and calculation complexity.