Contourlet-CNN for SAR Image Despeckling

A multiscale and multidirectional network named the Contourlet convolutional neural (CCNN) is proposed for synthetic aperture radar (SAR) image despeckling. SAR resolution not higher than that of optical images. If depth increased blindly, detail information flow will become quite weak, resulting in severe vanishing/exploding gradients. In this paper, a constructed, which single-stream structure layers replaced with multiple-stream to extract features properties, thus significantly improving despeckling performance. With help Contourlet, CCNN designed multiple independent subnetworks respectively capture abstract an certain frequency direction band. The can increase number by increasing subnetworks, makes only have enough features, but also overcome problem gradients caused deepening networks. Extensive quantitative qualitative evaluations real images show superiority our method over state-of-the-art speckle reduction method.