Synthetic aperture radar image change detection based on convolutional‐curvelet neural network and partial graph‐cut

Abstract

Synthetic aperture radar (SAR) images are widely applied in change detection tasks because of SAR's active imaging mechanism. However, SAR images suffer from speckle noise due to SAR reception coherence from distributed targets. This property of SAR increases the uncertainty of the image pixels, making it difficult to accurately detect the changed regions from the background. To address this issue, this letter proposes a novel SAR image change detection method based on a convolutional-curvelet neural network (CurveCNet) and partial graph-cut. The curvelet transform is introduced into the convolutional neural network to retain the structural information in the SAR images and to suppress the speckle noise. In addition, to combine the merits of various difference images, training samples are jointly constructed and fed into the neural network for it to learn the implicit common hierarchical features. Finally, the partial graph-cut approach is proposed to refine the pixel labels in the fuzzy region and inhibit outliers on the background for the resultant change map. Visual and quantitative results obtained on two real SAR image datasets have demonstrated the effectiveness and robustness of the proposed method.