LCDNet: Lightweight Change Detection Network With Dual-Attention Guidance and Multiscale Feature Fusion for Remote-Sensing Images

Deep-learning (DL)-based change detection (CD) techniques have recently become increasingly complex to produce more accurate detection results. However, the increase in complexity leads to reduced efficiency and limits the application of DL-based CD techniques in domains that require real-time performance. To this end, a lightweight CD network (LCDNet) is proposed to accurately recognize changes in remote-sensing (RS) image pairs while maintaining high efficiency. First, a focus module is utilized at the beginning of the encoding layer for the downsampling operation, which reduces the computation of the model and the loss of information. Then, a depthwise (DW) convolution-based efficient extraction block (EEB) is designed by stacking different sizes of convolution kernels for the effective extraction of change features under different receptive fields. Next, a dual-attention guidance module (DAGM) is designed to guide the encoder in processing and selectively aggregating information related to changes. Lastly, a multiscale feature fusion module (MFFM) with low parameters is proposed that combines feature maps of different scales to exploit their complementary information. Compared with other state-of-the-art (SOTA) methods, the proposed LCDNet only requires approximately 0.83 M Params, 2.03 G FLOPs, and 3.03 ms inference time (It) to remarkably surpass them in terms of accuracy. Moreover, compared with other dual-attention and multiscale fusion modules, the proposed DAGM and MFFM are more effective and efficient. The source code will be made available at https://github.com/sjl2023/LCDNet .