Model-Based Information Extraction From SAR Images Using Deep Learning

In this letter, a model-based approach to information extraction and despeckling is presented using a maximum a posteriori (MAP) estimate. An autobinomial model (ABM) is used as a prior and Nakagami model for a likelihood probability density function (pdf). ABM parameter estimation using the evidence maximization makes this method very computationally demanding and unuseful for practical applications. This letter proposes a deep convolutional neural network (CNN) for parameter estimation of ABM. A nine-layer CNN was used, consisting of convolutional, pooling, dropout, fully connected, and regression layers. The MAP estimation and CNN interactions were fused to obtain the best texture parameters with the highest evidence. Experimental results showed that the ABM using evidence maximization is 55 times slower than the ABM with CNN. Texture parameters are estimated better with the proposed technique because the estimated evidence obtained with the proposed method is much higher compared with the previous method. The improved parameter estimates subsequently used for MAP despeckling purposes provided favorable results.