Locating Building Change via Adaptive Frequency Enhancement

Building change localization for very-high-resolution image is important in accurately tracking urbanization. However, the increase in resolution inevitably enhances the complex and variable background interference, which affects the accurate identification of building targets. This paper seeks to explore a computational intelligence approach for tackling the above challenges. Specifically, a spatial frequency adaptive enhancement is presented that takes into account the spatial frequency difference of different land objects, so as to formulate the spatial frequency attention network. It can adaptively enhance the information that favors the detection of architectural changes and suppresses irrelevant background noise interference. The entire network is designed with a classic U-shaped architecture, and two attention schemes are specifically designed, including spatial frequency attention module to regulate spatial-frequency weights in each level of feature maps, and the triple attention gate to comprehensively integrate spatial, channel, and frequency information. Experiments on three popular real datasets show that our proposal is able to obtain advanced performance, and visualization of features indicates the positive effect of our spatial-frequency attention.