Fundus abnormalities caused by ocular and systemic diseases can be shown in retina. Therefore, the examination of the fundus is an applicable tool in diagnosing and monitoring both eye diseases and whole body conditions. Due to the existence of aberration, retina images which are acquired by the common camera can not be utilized to help ophthalmologist on focal location and diagnosis. Adaptive Optics (AO) technique solved this problem by revealing the smallest capillary vessel of the retina through compensating the aberration. However, the field of view is too small to provide enough information in the very small retina region revealed by one AO retina image. Moreover, it is inefficient and ineffective to mosaic artificially, and there are obvious marks in it . Hence, it is essential to carry out research on automatic mosaic technology for the AO retina images.The distributing of capillary vessel in retina is solid reticulation, hence, the images are not on the exactly same layer because of spherical surface of the funds and the aberrations of the eyes, which made the inconsistent details in the overlap area; meanwhile, the scale variety of the images is also existed because of the change of the focus; the characters of AO made the images bright in the middle area and dark in around area, reflectivity of retina is difference in different area, which made the images grey discordance; the images also have toothful rotation because the spherical surface of the funds ; so it is difficult to make the stitching of the high resolution human retina capillary vessel images by AO. Besides, 169 mosaicing multiple and multi-row images increase the difficulties severely. Then, the task of the project is the mosaic of the 169 AO retina multiple and multi-row images, which have scale variation, rotation, grey discordance and detail not corresponding.There are two parts in this thesis: one part is searching mosaic method of two images, the other part is analysing the distortion of multi-images mosaic, and a method is proposed. A robust stitching arithmetic based on feature points is presented. This method is based on the corner points which is detected by improved Harris arithmetic, then, initial point matching by RANSAC, accurate point matching by Mahalanobis distance, at last, affine alternate parameter are calculated, then the two images are stitched. However, the method mentioned aboved is produced seriouse problems when multi-images stitching is carrided through. All the images’ automatic stitching is not completed, the reasons are the imprecision of corner corresponding and transform parameter calculation. To the problem of imprecision of corner corresponding and transform parameter calculation, KLT arithmetic is adopted, after that, Fourierism’ similition as accurate point matching arithmetic, projective alternate parameters are calculated by LM iterative arithmetic. At last, the backward distortion and cap function are adopted to fuse the images,then the good results are obtained. The main task and achievement mentioned above are summarized as follows:(1) A corner detection arithmetic is proposed. Harris’ corner detection method is a widely used excellent detection method characteristed with high precision and low calculationn cost, but it is limited by the unstability for images after rotation and being added by noise. We propose a new corner detection method which is propitious to detection the accurate the corner of the OA retina image.(2) A corner points matching arithmetic is presented. For the characteristics of retina images, such as scale variety, noising, grey discordance, detail not corresponding, etc, a corner points matching arithmetic is presented, initial point matching by RANSAC, accurate point matching by Mahalanobis Distance distance, which can get accurate corner point matching by a few calculation.(3) A method is presented for multi-images stitching. It is based on KLT corner detection, using Fourierism’ similition as accurate point matching arithmetic, and projective alternate parameter are calculated by LM iterative arithmetic, and this method is implemented.(4) The problem of the confusing of multi-images is solved. Firstly, the problem is predigested to the confusing of round images overlaped with each other, a settled method is proposed, and good result is achieved.The well mosaicing of large vision and high resolution human fundus images provide true and reliable evidence for early diagnosis of disease. This technique can not only prevent many blinding disease, but also have significance for precausion of many systemic and nervous diseases. From the point of image mosaicing technique, it is also new to solve the problem of mosaicing multiple and multi-row images with rotation, transformation, inconsistent intensity and details. This not only solves the problem of accumulative error of mosaicing multiple and multi-row images, but also solves the problem of fusing multiple and multi-row images. This algorithm is applicable to the mosaicing and registration of multiple and multi-row images which have small scale variance, rotation, unpredictable translation and lower correspondence between images.