Research Of Image Feature Detection And Matching Based On Phase Information

Image matching technology is the core of digital photogrammetry. Image feature detection is one of the key technologies of image matching, registration, segmentation and object identification, high-precision feature extraction and positioning is a base to extract image features automatically from remote sensing images, and the technology is of great significance to rapidly update the GIS data. Aiming at those problems occurring to both feature detection and matching methods based on spatial domain analysis, this thesis develops the studies on feature detection and image matching based on frequency domain analysis. For feature detection, traditional methods are easy to be affected by image brightness, contrast, threshold selection and so on. So, this thesis introduces phase information in frequency domain to detect image features. Phase information plays a decisive role on the system of human visual perception, it contains much information. Phase congruency is a new feature and has shown a potential advantage in feature detection. In image matching, it is a formidable problem how to determine an initial disparity value. So, this thesis develops a matching method based on phase correlation and then proposes a layered FM phase correlation algorithm by combining the idea of layering and blocking image circularly to match images. This method can automatically acquire an initial disparity value of image matching. In image matching, a stepwise block partitioning phase correlation approach is presented in order to obtain dense disparity of image. The approach shows good performance in matching images of flat area, mountain areas with large elevation change and deformation and forest area with broken textures.The main contents of the thesis are as follows:1) The developing situations of image feature detection and matching methods are presented, and advantages and disadvantages of those traditional methods are discussed. On the basis, the thesis develops experiments and analysis on typical feature detection algorithms based on spatial domain analysis under the different parameters.2) Image feature detection theories and methods based on phase information are stated in detail. The focuses are on studying the relationship between phase congruency and local energy model, the measurement of phase congruency, the selection of filter and feature detection of two-dimensional image etc. More importantly, the thesis proposes a novel method of power line feature detection from images by combining phase congruency with the one-dimensional directional filtering. 3) The comprehensive analysis of phase information, direction and phase congruency strength of detected features is firstly carried out; then a method to distinguish the different type of image features is studied. As phase congruency model is built according to the assumption that human visual system can sense that features are considered as the points whose phase of Fourier components are the most consistent, it is not to detect a specific feature. In the ideal case, those points with maximum phase consistency can be considered as the feature points. In practice, image features are usually divided into different types, such as:corners and lines, both which are different in dimension, and edges (i.e, the boundary of two different texture region) which is corresponding to step signal different from line, which is corresponding to the impulse signal representing an drastic change of brightness in flat region. In this thesis, features of one-dimensional signal in all directions are detected, then integrated the results of every direction using moment analysis so as to corners and lines are distinguished and detected. In addition, a method to distinguish the type of edge using phase information was proposed after an analysis of phase difference of various edges.4) Compared to the image amplitude, phase contains more information. Amplitude information can be used to determine changes of gray scale while phase information can be used to determine the structure and location information. Because of the importance of phase information, a phase-related technology is developed. Specifically, the thesis studies a phase correlation approach by combining Log-polar transformation-based and Fourier-Mellin domain-based phase correlation techniques in order to solve the issue of stereo image correlation in the context of rotation and scaling transforms. Both the precision and calculation speed can’t be increased simultaneously when the parameters of rotation and scale are resolved by phase correlation based on Log-polar transformation or Fourier-Mellin transformation. So, an improved phase correlation approach—logarithmic phase correlation algorithm by progressively refining is proposed. The approach can achieve a sub-pixel accuracy and show well in image registration, also can increase the computation speed regardless of the size of image. Thus the overall performance of the algorithm is improved and the scope of application was broadened.5) How to get an initial disparity value is a key issue in stereo image matching. After a summary of traditional methods to get the initial disparity value of images were summarized and analyzed, it is found that there are many problems of how to determine an initial disparity value, such as:low automation, poor accuracy and low computation efficiency and so on. For these problems, the thesis proposes a method called as layered FM phase correlation to obtain initial disparity value. The method can efficiently and automatically obtain a more accurate initial disparity value without additional information, even if stereo image pairs are blocked by objects, it can obtain an initial disparity value with high precision.6) Combined with the idea of image blocking, a stepwise partitioning phase correlation approach is presented. Experiments show that the method can perform well in matching images in flat area, in mountain areas with large elevation change and local deformation, and also in forest area with broken textures.