Research On Image Registration Algorithm Based On Local Feature And Its Applications

Image registration is a key technique in pattern recognition and image processing,and it is widely applied in many application areas such as computer vision,remote sensing,image fusion,mage super-resolution reconstruction and medical image processing. The requirement for improving the accuracy and efficiency of image registration methods increases along with the development of their applications. There are two main categories of image registration method: area-based method and feature-based method. The feature-based image registration methods utilize features instead of whole image to obtain higher accuracy and efficiency. This paper mainly concerns about the feature-based image registration methods, which is a popular topic in current image processing area. The paper studies about coordinate transformation estimation and feature acquisition techniques under complex imaging conditions. Several novel registration methods are proposed based on our research and their advantages are also verified in this paper.The research and innovative achievements include several aspects as follows:1. To improve the CPD algorithm(Coherent Point Drift),the P-CPD algorithm is proposed. P-CPD algorithm replace the displacement function with projection parameter function,which provides more degrees of freedom for point coordinate transformation. The projective transformation matrix of each image pixel can be easily obtained by using the projection parameter function. Experiment results show that the P-CPD algorithm can achieve lower registration errors under disturbance of noise points and outliers. F-CPD algorithm is proposed to improve the P-CPD algorithm by taking advantage of local image information. Local features are introduced to the Gaussian mixture model of F-CPD algorithm to accelerate convergence and improve accuracy. Experiment results show the improvement of F-CPD algorithm in terms of accuracy and efficiency.2. Image coordinate transformation parameters can be estimated by utilizing the image point matches. MDLT algorithm(Moving Direct Linear Transformation)is a projective estimation technique,which is fast but not robust to the outlier matches. To overcome this disadvantage,VF-MDLT algorithm is proposed. Instead of computing transformation parameters for each image pixel,VF-MDLT algorithm estimates projective vector field toobtain projective matrix of each pixel. Projective vector field provides better capacity of resisting outlier disturbance. VF-MDLT algorithm is also used in image stitching to give better alignment in overlapping area,together with SPHP algorithm which preserves the perspective of non-overlapping image region. Experimental results show that the VF-MDLT algorithm can provide better alignment with the existence of outliers. And the proposed image stitching technique can manipulate stitching images with less ghost effects in overlapping area and less projective distortion in non-overlapping area.3. According the characteristic of multi-focus image,we propose the SLS-SP descriptor and corresponding matching method,which can provide reliable image matches to carry on image registration. A novel focus stacking method is also proposed to merge the registered multi-focus image sequence to extend the depth of field(DoF)of synthetic image. We approach the image fusing as label assignment problem,which is suited to a Markov Random Field. VF-MDLT algorithm is chosen as the registration method. Experimental results demonstrate the significant improvement of matching accuracy by using SLS-SP descriptors in blurred image. Our focus stacking method together with our image registration algorithm can extend DoF while retaining the visual verisimilitude of the scene.4. A fast and accurate matching method for estimating dense pixel correspondences across scenes is proposed. It is a challenging problem to estimate dense pixel correspondences between images depicting different scenes or instances of the same object category. While most such matching methods rely on hand-crafted features such as SIFT,we learn features from unlabeled image patches using unsupervised learning technique. The learned features are seamlessly embedded into a multi-layer matching framework. We experimentally demonstrate that the learned features,together with our matching model,outperforms state-of-the-art methods both in terms of accuracy and computation efficiency.Furthermore, the performance of a few different dictionary learning and feature encoding methods is evaluated in the proposed framework,and the impact of dictionary learning and feature encoding with respect to the final matching performance is analyzed.