Enhanced Self-Attention Network for Remote Sensing Building Change Detection

The self-attention mechanism can break the limitation of receptive field, model in a global scope, and extract information efficiently. In this work, we propose lightweight remote sensing building change detection (ESACD). encoder, use enhanced layer, CoT instead normal convolution operation. layer fuses dynamic context with static context. Compared ordinary convolutional strategy fully mine local features between input keys to dynamically enhance feature representation. Subsequently, dual attention further low-frequency high-frequency images semantic interest model. Dual consists HiLo Tokenizer mechanism. extracts through two branches. branch, nonoverlapping windows are applied for self-attention. average pooling is first performed on before After tokens that interested in, it encodes its and, then, converts into pixel-level features. realizes efficient extraction enhances representation ability Finally, fuse fluidity improve accuracy. Through our experiments datasets, ESACD has better performance than other state-of-the-art methods while maintaining fewer parameters.