The main goals of this work are to propose a more automatic and efficient algorithm to replace regions of clouds and shadows in satellite images as well as an index of reliability that is previously applied to each image, in order to measure the feasibility of the estimation of the regions covered by atmospheric components using that algorithm. The motivation comes from the problems caused by these atmospheric elements, among them: to impede the identification of objects of the image, to make the urban and environmental monitoring more difficult, and to interfere in crucial stages of digital image processing to extract information for the user, such as segmentation and classification. Through a hybrid approach is proposed a method for decomposing regions using a median non-linear low-pass filter, in order to map the regions of structure (homogeneous) and texture (heterogeneous) in the image. In these areas was applied restoration methods Inpainting by Smoothing based on Discrete Cosine Transform (DCT), and Exemplar-Based Texture Synthesis, respectively. It's important to note that the techniques have been modified to be able to work with images obtained through of satellite sensors with peculiar features such as large size and/or high spectral variation. Regarding to the reliability index, it aims to analyze the image that contains atmospheric interference and hence estimate how much reliable will be the redefinition, based on the percentage of cloud cover over the regions of texture and structure. This index is composed by combining the result of supervised and unsupervised algorithms involving three metrics: Average of Accuracy Global, Measure Of Structural Similarity (SSIM) and Average of Pixels Confidence. Finally, it was verified the effectiveness of these methods through a quantitative assessment (provided by the index) and qualitative (the images resulting from processing), showing the possible application of the techniques to solve the problems that motivated this work.
|
