Rethinking Building Change Detection: Dual-Frequency Learnable Visual Encoder With Multiscale Integration Network

Remote sensing (RS) image change detection (CD) methods based on deep learning, such as convolutional neural networks (CNNs) and transformers, are still spatial domain-based image processing methods by nature, and their detection accuracy is strongly affected by chromatic aberration due to imaging time, shadows caused by lighting conditions, and object confusion and other disturbances. In this study, we revisit CD from a signal processing perspective, framing it as the task of consistency detection of the distributional features of two 2-D signals. We aim to extract the primary components of the two signals while suppressing interfering noises. To address this, we propose a novel CD method called DFNet, which leverages a dual-frequency learnable encoder. First, we construct a dual-frequency feature encoder Siamese framework to capture local high-frequency signals and global low-frequency signals using CNN and attention mechanisms after dividing the input RS image signals into two channels. Second, we introduce the frequency explicit visual center module as a part of the multifrequency-domain dense interaction (MFDDI) module at the decoder stage, allowing long-distance dependency to be established between high–low frequency components in the same layer as well as signal aggregation in regions of small edge variations. In addition, the MFDDI module adopts a layer-by-layer interactive fusion approach to synthesize discriminative information in a wide frequency-domain range, enhancing the characterization capability of frequency-domain signals. We conduct comparison experiments with the current mainstream methods on the land cover dataset SYSU-CD and two building datasets, LEVIR-CD and WHU-CD, and the results show that our method is not only resistant to interference but also outperforms all the comparison methods.