Mapas de entropia e esperança logarítmica em processamento de imagens SAR

Abstract

Synthetic Aperture Radar (SAR) has been used for many years in activities related to global earth monitoring. There are many advantages of using SAR systems when compared to standard optical ones, since they operate independently of the weather and of the presence of sunlight conditions. However, due to the coherent illumination of these systems, SAR images suffer strong contamination by speckle noise. In fact, this interference significantly degrades the quality of SAR images, which leads to difficulty in image processing tasks, making it difficult to use traditional methods, such as image thresholding. Given the importance of synthetic SAR images in the algorithm design and performance analysis for image processing, we first evaluated the generation of synthetic SAR data, modeled by the G 0 A distribution, using two approaches. Our results demonstrated that both approaches perform quite similar to each other, however they differ remarkably in the processing time. Actually, the indirect method is 10x faster than the direct approach to generate synthetic SAR data. In this thesis we have also proposed two methodologies that make use of the map of entropy and the map of logarithmic expectation, taking into account that the SAR data follow the model G 0 A for amplitude, or G 0 I for intensity. To evaluate the performance of the proposed methodologies, we perform segmentation experiments, measuring the Error of Segmentation, the Cross-Region Fitting index, the rates of false positives and negatives, besides indexes of similarity. Tests performed on synthetic and actual SAR images generally indicated that both proposed maps improved segmentation results, regardless of the increase of the number of looks. Regarding the computational time, both maps presented satisfactory results, with emphasis on the map of logarithmic expectation, which obtained the least computational time. In order to consolidate the efficiency gain with the use of these maps in SAR image processing, we performed image retrieval experiments. The results of these experiments showed a positive predictive value (PPV) above 90 % and a mean of the mean accuracy (MAP) scores above 99 % indicating that both approaches can also be used in content-based applications.