Enhancing VMamba for change detection via lightweight feature interaction and selection

Change detection aims to identify areas or objects of interest that have changed between bi-temporal images, which is a hot topic in the remote seeing and pattern recognition community. Recently, the visual state space model (VMamba) has demonstrated impressive and efficient results compared to previous methods based on convolutional neural networks and Transformers. However, existing VMamba-based models ignore the domain gap in bi-temporal images, which limits the potential of VMamba for change detection. To this end, a Siamese VMamba with Interaction (VMI-CD) model is developed. Specifically, a parameter-free bi-temporal feature interaction module (BFIM) is proposed, which is custom-designed for VMamba during the feature encoding stage to enhance the perception of the model in domain differences. Besides, a channel-spatial selection module (CS2M) is designed to modulate bi-temporal features, which aims to facilitate the generation of discriminative change representations in difference extraction. Thanks to the lightweight design, only a small number of parameters are added with the introduction of BFIM and CS2M. Experimental results on five remote sensing image change detection datasets with different tasks, MCLC-CD and JL1-CD, which contain multiple change types, LEVIR-CD+ and WHU-CD for building change detection, and SYSU-CD, a category-agnostic binary change detection dataset, demonstrate that VMI-CD surpasses previous state-of-the-art approaches. The code will be available at https://github.com/ptdoge/VMI-CD.