Abstract: Aiming at the problem of low recognition rate and poor robustness caused by weak echo signal and multi-scale feature coupling in ultra-wideband (UWB) radar gesture recognition, a recognition method combining multi-scale feature decoupling and dual-channel co-attention was proposed. The clutter suppression process combining moving target indication (MTI) and robust principal component analysis (RPCA) is used to achieve the hierarchical suppression of static and low-velocity clutter, which effectively improves the signal signal-to-noise ratio, and then generates a high-resolution time-frequency image by short-time Fourier transform (STFT). The multi-scale decoupling feature network (MSDFNet) is proposed, which explicitly separates global motion and local micro-motion features by parallel design of large-scale depthwise separable convolution and small-scale standard convolution to realize the decoupling of multi-scale features. A dual-channel attention mechanism is introduced to adaptively enhance key features. The experimental results show that the proposed clutter suppression method improves the average signal-to-noise ratio by 10.8 dB, and the clutter suppression rate reaches 86.4%. On the basis of the preprocessing, the overall recognition rates of the model on the self-built dataset and the public dataset are 97.8% and 98.1% respectively. The proposed method significantly improves the accuracy and robustness of gesture recognition, and provides an efficient and reliable solution for non-contact gesture interaction.