Style Adaptation and Deformable Convolution for Change Detection Networks

Remote sensing image change detection plays an important role in the fields of urban change detection, environmental protection and geologic hazard identification. Change targets are mostly irregular shapes, and it is a challenge to extract the change characteristics of remote sensing images using ordinary convolutional methods because the perceptual field of convolution is fixed. In this paper, a style adaptation and deformable convolution network(SDNet) is proposed and use a deformable convolutional network with residual links in SDNet. Global context and change characteristics details can be captured in a much shallower way. In addition, a special detection module is proposed for comparing multilevel features extracted from two branches, where the features of each branch are shared. In order to reduce the background and illumination of the diachronic images, a style adaptation technique is used to make the two images similar in style. The effectiveness of our proposed approach has been validated through experiments conducted on two public datasets.