Abordagem estocástica para recuperação de imagens SAR baseada em conteúdo

Abstract

This thesis proposes a Content-Based Image Retrieval (CBIR) system using stochastic distance for Synthetic Aperture Radar (SAR) images. The methodology considers the $G_{I}^0$ distribution to describe \gls{SAR} data in intensity. Based on this premise, three essential steps were established to retrieve SAR images as well as to evaluate the results: i) Estimation of the roughness ($\alpha$) and scale ($\gamma$) parameters of the distribution $G_{I} ^0$, i.e., the image feature extraction step. To estimate the parameters, two estimation methods were tested: the maximum likelihood and a fast approach of the logcumulant method; ii) The use of the triangular stochastic distance as a similarity measure. The method evaluates the similarity between a query image and the other images in the database to perform the content-based image retrieval. The stochastic distance identifies the most similar regions according to the image characteristics that are the estimated parameters of the data model; iii) Evaluation of our proposal applying the \gls{MAP} measure and considering clippings of an image from each radar sensor, that is, UAVSAR, OrbiSAR-2 and ALOS PALSAR. The results with the CBIR-SAR method were obtained for synthetic images that achieved the highest MAP value, recovering extremely heterogeneous regions. Using real SAR images, the CBIR-SAR method obtained MAP values above 0.833, for all polarization channels for forest image samples (UAVSAR) and urban areas (ORBISAR). Our results confirmed that the proposed method is sensitive to the texture present in the image and, and thus it relies on the quality of the estimated parameters that are inputs of the stochastic distance to perform the effective image retrieval.