Simultaneous-source acquisition technology can improve acquisition efficiency and reduce exploration costs, and its core is the deblending method. Traditional deep learning-based deblending methods are prone to losing detailed boundary information during the down-sampling process, thereby reducing the signal-to-noise ratio of the separation results. This paper proposes a multi-scale UNetPlus (MsUNetPlus) deblending method. Based on the UNetPlus network, an independent "left leg" path is introduced in the encoder to extract shallow features of the input image at different resolutions, which are then closely fused with the main encoder path. This helps supplement edge details at deeper levels and alleviates the edge blurring caused by down-sampling. Separation tests on physical simulation data show that the proposed method achieves higher separation accuracy compared to the UNet, MsUNet, and UNetPlus methods, and can better preserve image detail information. Finally, the proposed method was applied to field data and achieved satisfactory separation results.