R&D-Net: Integration of Registration-Net and Detection-Net for Identifying Building Changes in High Spatial-Resolution Remote Sensing Images

In the detection of building changes in high spatial-resolution remote sensing images, both the distribution position and surface characteristics of the same objects are probably different under different imaging phases, which potentially causes high false positives. In order to improve the detection accuracy of building changes, an integration network, named R&D net is proposed in this article, which comprises a registration network (R-net) followed by a change detection network (D-net). In R-net, two different phase images are accepted as inputs, corner points and their descriptors are generated to spatially align those images. After that, the spatially aligned images are fed into the D-net, and building images are detected accordingly. In this article, a multiview automatic labeling method is proposed to obtain labeling corner points. A new dataset containing 5104 image pairs is established. Experimental results demonstrate that the R-net can extract robust invariant features, and then improve registration accuracy under circumstances with obvious changes of surface feature, which is a base of D-net. Uniting pyramid pooling structure with a focal loss function in D-net, both leaky and wrong segmentations can be dramatically improved under complex scenes with many interferences. When compared with baseline methods on different high-resolution remote sensing scenes, the proposed method achieves better performance and more accurate detection results of building changes.