Research On Key Techniques Of Storage And Management For Large Scale Images

As the technology of remote sensing develops rapidly, it has become a very important subject to use large scale images effectively. This study is based on an advanced researching project. The goal of the project is to give support for operation, using information got by satellite. For this purpose, the author did some researches on some key techniques for storage and management of large scale image, including the following ones.In this study, it is a primary task to build a storing model for large scale iamges. Image pyramid is an important kind of multi-resolution structure. In this dissertation, a class of modified Reduced-sum pyramid is proposed, in order to solve the problems of"precision truncation"and"data expansion". As long as a proposed general formula is fulfilled, the resulting pyramids are free of the above two problems. Furtherly, three practical models are also provided. Both theoretical analysis and experiments show that this type of pyramid has excellent performance of storage and visual quality, while free of the above problems.Although efficient storage models are built, it is still necessary to perform data compression for them. Some researches show that the distribution of pixels of an image has effect on the performance of compression algorithms. In this dissertation, a lossless compression algorithm and a near lossless one are proposed, respectively, based on this idea. In order to improve the performance of LZW algorithm, the proposed methods change the distribution of pixels using pre-processing steps. Experiments show that they both gain general improvement on compression ratio.Various remote sensing images and maps are involved in this study, with different types of projection. Therefore, it is a frequent task to perform projection transformation. A method is proposed in this dissertation to solve this problem, combining analytical solution and numerical solution. It divides all the pixels into four ranks and handles these ranks using different approaches. Accuracy and speed of transformation can be adjusted flexibly by choosing parameters of ranking. Experiments prove this character and show that fairly high accuracy can be got in short time.As a basal part of the whole project, this study is demanded to provide necessary data for other systems. One of the tasks is to extract vector data from raster images. Image classification is an important approach for transforming raster data to vector data. In this dissertation, K-means algorithm is modified, employing spatial context information. The modified algorithm keeps the strongpoints that K-means has, and it depends not only on spectral information but also spatial information. Experiments on many real images prove this character. It can also be seen that the proposed algorithm is far less sensitive to noises, compared with K-means and ISODATA. This dissertation has very close relationship with the project. Therefore, the author has applied most of the above techniques to the system, and got satisfying results.