CAIFNet: Capturing Amplitude-Invariant Features for Remote Sensing Image Change Detection

Change detection (CD) is a critical task in remote sensing (RS) image analysis. Recent deep learning networks for CD focus on identifying changes after mining the features of bi-temporal images separately. However, light differences in bi-temporal images lead to the networks extracting different features from the identical objects, which may cause pseudo-changes. From the Fourier transform perspective, an image can be decomposed into amplitude and phase, where the amplitude contains most of the light information and the phase is relevant to structure information. Therefore, amplitude-invariant features of the identical objects in different light conditions are roughly the same, which are pivotal to identify real and fake changes between bi-temporal images. In this article, we propose a capturing amplitude-invariant features network (CAIFNet), which reduces dependence on amplitude and captures diverse amplitude-invariant features. Firstly, we build an amplitude pre-processing module (APM) to provide diverse processed images by randomly mixing the amplitudes of the input images with the amplitudes of the reference images and keeping the phases of the input images constant. Secondly, a quadruple-stream encoder is proposed to capture amplitude-invariant features. Specifically, it is forced to learn and capture amplitude-invariant local details and amplitude-invariant contextual semantics based on the diverse processed images under CD task-oriented constraint, both reciprocate each other to become more accurate by local attention guide strategy (LGS). Moreover, a difference enhancement module (DEM) is designed in the quadruple-stream encoder to enhance the difference features. Thirdly, a bi-stream decoder decodes the captured amplitude-invariant features in main and boundary difference perspectives, enhancing main body and boundary details of the objects in the change maps, respectively. Finally, a spatial embedded module (SEM) allows the main and boundary difference features to be embedded into each other, obtaining more complete change maps. On three remote sensing change detection (RSCD) datasets, CAIFNet achieves better transferability and results compared to state-of-the-art methods. The source code is available at https://github.com/yihui1230/CAIFNet.