Geometry-Feature-Registration-Based Robust Image Mosaic Algorithm For Images

Image Mosaic technique is an important subject of research in computer vision, pattern recognition, medical science and other fields. And it is one of the key technologies of image processing. The so-called image mosaic is such a technology that is used to seamlessly join multiple overlapped images into a wide horizon image. Image Mosaic includes two key technologies, namely image matching and image fusion. Because of the uneven illumination of local field images, the difference between illumination of the images, and the geometry deformation between images, it is difficult to implement perfect matching. The technical bottleneck is the accuracy of sub-pixel registration and the seamless integration on both sides of the overlapped area of images. In this paper, the illumination-robust mosaic technology based on precise image matching is the main work. Through the study on representation and matching computation of geometry feature points and line style feature, we further study the feature matching algorithm which has the feature of rotation invariant, and is robust to illumination changes. Then we will design image matching algorithm with high performance and robust image mosaic method.The main research contents of this paper are as follows:(1)A kind of fast image registration framework based on Shi-Tomasi corner matching and corner feature spectral clustering is proposed in this paper. This method could construct adaptive image projection transformation based on an8-parameters perspective transformation model. And this method could provide an algorithm for corner features between images from coarse matching to accurate matching. Through combining projection parameter fitting of cross-correlation based on coarse matching and improved random sample consistent accurate matching, the image accurate matching and mosaic are implemented. Experimental results show that the method not only improves the efficiency of the algorithm, but also enhances the image registration and stitching in light of robustness.(2)A simple and effective image registration algorithm based on rotation invariant HOG feature is proposed in this paper. HOG feature is a feature which is based on gradient direction of statistics. Gradient direction is obtained by direct corresponding pixel gradient. However, when the registration image rotation happens, the effect obtained through single gradient direction information is not ideal. In order to solve this problem, the paper proposes a rotation invariant HOG feature by increasing the description number of gradient direction. The first step is to use Harris corner detection algorithm to evaluate some rapid methods of extracting key point. The second step is to describe the key characteristics of rotation invariant HOG feature. Finally, fulfill the task of image registration by point pattern matching. Compared with similar methods, this method has a better robustness in rotation changes.(3)A registration model about description of the image structure characteristics is proposed. The current mainstream method of image registration misses the feature information based on single point. This kind of method lost the image structure information imperceptibly. In other words, this method ignored the key point between the neighbor information. In order to compensate for this weakness, this paper proposes a directed line segment based on describing the image of SIFT features. Firstly, we use the method of Harris corner detection to detect the key points. Secondly, we describe the SIFT feature points between key points to constitutes the characteristic matrix and describes the directed line segment. At last, we use the Ransac method to get the point matching to complete the task of accurate image registration. The experimental results show that this method provides superior performance when the image is suffering from distortion, deformation and other problems, especially when the scale of matching image changes.(4)In order to solve the problem that suture line may be close to the boundaries of transition, the transition region equalization techniques is proposed to extract an optimal suture line from two overlapping boundaries as far as possible. Moreover, in order to overcome the disadvantage that the splicing trace cannot be eliminated sometime when single Gaussian function is used to implement weighted band-pass image fusion for multi-resolution fusion, we propose an adaptive weighted fusion algorithm based on the parameters of the cosine function However weighted fusion easily weakened the information of edges. To avoid this problem, the information of edge is referenced when fusing. Experiment results show that the transition area after using the proposed method has better visual effects.