DCAT: Dual Cross-Attention-Based Transformer for Change Detection

Abstract

Several transformer-based methods for change detection (CD) in remote sensing images have been proposed, with Siamese-based methods showing promising results due to their two-stream feature extraction structure. However, these methods ignore the potential of the cross-attention mechanism to improve change feature discrimination and thus, may limit the final performance. Additionally, using either high-frequency-like fast change or low-frequency-like slow change alone may not effectively represent complex bi-temporal features. Given these limitations, we have developed a new approach that utilizes the dual cross-attention-transformer (DCAT) method. This method mimics the visual change observation procedure of human beings and interacts with and merges bi-temporal features. Unlike traditional Siamese-based CD frameworks, the proposed method extracts multi-scale features and models patch-wise change relationships by connecting a series of hierarchically structured dual cross-attention blocks (DCAB). DCAB is based on a hybrid dual branch mixer that combines convolution and transformer to extract and fuse local and global features. It calculates two types of cross-attention features to effectively learn comprehensive cues with both low- and high-frequency information input from paired CD images. This helps enhance discrimination between the changed and unchanged regions during feature extraction. The feature pyramid fusion network is more lightweight than the encoder and produces powerful multi-scale change representations by aggregating features from different layers. Experiments on four CD datasets demonstrate the advantages of DCAT architecture over other state-of-the-art methods.

LRDE-Net: Large Receptive Field and Image Difference Enhancement Network for Remote Sensing Images Change Detection

“…The structure of this network is complicated, and its computational efficiency is low. Zhou et al [31] introduced a dual cross-attention transformer (DCAT) network. This network is designed to extract both low-frequency and high-frequency information from input images through the computation of two distinct types of cross-attention features.…”

DTT-CGINet: A Dual Temporal Transformer Network with Multi-Scale Contour-Guided Graph Interaction for Change Detection

“…Immediately after, we use the DTT decoder to reweight the original features based on the generated tokens to obtain the refined features considering the dual-temporal contextual relationships. While some works, such as [21][22][23], employ transformers based on cross-attention for change detection, their proposed cross-attention merely involves the straightforward calculation of attention matrices using the query (Q) from another temporal phase and the key (K) from the current temporal phase. This approach fails to adequately model the non-local structural relationships depicted in Figure 1.…”

Enhanced Wind Field Spatial Downscaling Method Using UNET Architecture and Dual Cross-Attention Mechanism

“…For instance, simple skip connections in encoder and decoder features can lead to semantic gaps. Inspired by achievements in the field of medical image segmentation [27][28][29], this paper introduces a Dual Cross-Attention module (DCA) based on the UNET architecture, incorporating Channel Cross-Attention (CCA) and Spatial Cross-Attention (SCA) mechanisms. This DCA module adaptively captures channel and spatial dependencies between multi-scale encoder features in sequence to address the semantic gaps between encoder and decoder features in the UNET architecture.…”