The objective of this project is to apply Convolutional Neural Networks (CNN) to optimization based Synthetic Aperture Radar (SAR) imaging to learn good parameter choices that enhance desirable features while suppressing undesirable properties of the SAR images of tactical ground vehicles. Specifically, the combined CNN and imaging algorithm are designed to achieve a sharper shadow feature and to reduce the speckle noise in the SAR image. A convolutional neural network is a subset of machine learning based on artificial deep neural networks which are inspired by the biology of the brain. The significance of deep networks over previous work in artificial neural networks is the use of large number of layers as compared to the standard 3 layer network. The CNN is trained in a supervised regression setting where a human operator provides the truth which pairs desired parameter values with each image in the training set. These learned parameters are then predicted by the CNN for each test image and the parameters are fed to an optimization based image formation algorithm that is regularized by an edge detection term based on polynomial annihilation.The optimization is then solved using an alternating minimization approach. Several experiments are run comparing different networks, different learning algorithms (adagrad, adadelta and rmsprop), and different input normalization techniques (variable scaling and z-score). Comparisons between the human optimized images and the learned CNN images were compared with subjective visual comparisons and with objective measures including Mean Squared Error (MSE) and the correlation coefficient of both the predicted image and its inverted amplitude image. Using this approach, the adagrad normalized input network performed the best with a MSE of 0.0013 and a correlation coefficient of 0.9667 (un-inverted image) and 0.9772(inverted image).
|