Synthetic aperture radar image despeckling based on modified convolution neural network

In remote sensing applications, the extraction of relevant information on speckle noise-affected synthetic aperture radar (SAR) images becomes problematic. Thus, it is important to improve the quality on those images to get the correct information. To meet this requirement, a Modified Convolutional Neural Network (M-CNN) algorithm is proposed, in which despeckling is based on deep learning method which uses CNN with 12 layers. Our network also uses dilated convolution to enlarge receptive field and a leaky rectified linear unit (LReLu) activation function together with batch normalization (BN) and residual learning strategy. Skip connection is incorporated to avoid quality degradation in the input image, and quality loss is estimated as the total of Euclidean and total variation (TV) loss and mean squared error (MSE). Experimental results on real and synthetic images proved that our method outperformed state-of-the-art methods.