| The rapid development of computer, communication and remote sensing techniques has brought new opportunities for the remote sensing image processing since the 1980's. By virtue of its wide area coverage and various spatial resolution, remote sensing image has been widely used in different fields, including land resource investigation, basic geographic information acquisition, city planning, agricultural production, oceanographic monitor, meteorological observation and astronomical discovery. However, except active radar photography can effectively avoid the weather disturbance, almost all remote sensing images obtained in the passive optical means are subject to different levels of clouds, mist and brume's influence, and this interference has been the major obstacles for the widespread use of remote sensing image, especially for the mapping production based on remote sensing image, prolonging the production time and lower productivity.The quality and utilization of remote sensing image can be greatly improved when the clearness processing is performed based on single image and information compensation related to the thick cloud area is achieved based on auxiliary image. Aiming at the optical image commonly used in the earth observation, with the latest theories of photogrammetry, computer vision, image processing and graph theory, this thesis made an intensive and deep study on the theories and approaches of haze removal and information compensation related to the cirrus area, designed and completed a practical application software which has successfully been applied to the image processing on specific land areas and island areas.The main contents and innovations in this thesis can be summarized as follows.1. The electromagnetic spectrum features, textural features, spatial distribution features and correlation features in degraded cloudy images are summarized and analyzed. The haze image is made up of two signals, reflected electromagnetic wave signal from haze and ground, transmission electromagnetic wave from haze, and that degraded image with thick cloud only includes cloud reflected signal. These different degradation models determine two kinds of image processing methods. Above-analysis offers the essential theoretical supports for the study of subsequent algorithms.2. The main principles and fundamentals of classical haze removal algorithms are summarized and analyzed, which include histogram enhancement, homomorphic filtering, wavelet transformation, and dark channel prior dehazing. Then we conducted in-depth study of wavelet transform. There are many kinds of wavelet functions and transformations, which bring much trouble to the study of cloud removing in RS image. In this thesis, various types of wavelet base functions, decomposition layers and kernel methods are conducted and tested. The experimental results show that, db4 and sym6 wavelet functions with approximately symmetry provide better results in which the Mosaic Show has been eliminated entirely and no displacements at object's edges occurs. DSWT is applicable to three layers decompositions, and wavelet packets is suitable for two layers decompositions. Just one layer decomposition can produce the ideal results when the lifting wavelet is used. The lifting wavelet transform and stationary wavelet transform are fit for cloud removing process. From the above comprehensive analysis, the lifting wavelet transform using the db4 wavelet function and one layer decomposition is considered to be the optimal selection for haze removal, and the stationary wavelet transform using the db4 wavelet function and with three layer decompositions is the secondary selection.3. A clearness improvement method for image with non-uniform cloud and mist distribution is proposed. In this method, the density of cloud and mist is automatically evaluated based on the low pass characteristics of Gaussian Filter. Then the degraded equation of cloudy and misty image is improved, which avoid"under-processing"and"over-processing"in the clearness improvement. Finally, two images with different resolutions are used to verify the experiment. Experimental results indicate that the clearness improvement method presented in this thesis for the image with non-uniform cloud and mist distribution takes advantage over dark channel prior method in each evaluating indicator.4. Direct-replacing method is a simple and effective cloud removing method for the remote sensing images. This method substitutes the cloudy pixels with the corresponding pixels on the auxiliary image. But it inclines to cause obvious seam lines between substituted cloud area and original image. In this thesis, the replace area is skillfully selected firstly. Then combined with histogram theory and mathematical morphology theory, the scheme for the seamless information compensation related to the cloud area is discussed, which achieves good effects in the practical production application.5. By applying Min-cut/Max-flow of graph theory to the image dodging algorithm, water area and terrestrial area can be processes separately, which effectively avoids the water specular reflection effect. Experiments show that, compared with the algorithms based on Split-Merge and gradient field, the dodging algorithm based on the graph theory can achieve a more satisfactory results.6. The above algorithms are applied to the remote sensing image processing in islands areas and the southwest of china areas. Moreover, the rationality and effectiveness of the algorithms are verified in real projects. Key words: Information Compensation, Homomorphic Filtering, Wavelet Transformation, Non-uniform Cloud, Image Registration, Image Fusion, Graph Cut, Image Dodging Processing...
|
PDF Full Text Request