Recently, deep learning-based blind watermarking schemes have attracted increasing attention due to their effectiveness in copyright protection and traceability. Although recent schemes have demonstrated promising performance, they still struggle to achieve a balanced trade-off among imperceptibility, robustness, and model efficiency. Some schemes rely on local feature modeling, which limits their ability to capture global image semantics and leads to poor robustness under complex distortions. Transformer-based models employ high-complexity self-attention mechanisms that significantly increase computational cost and hinder deployment in resource-constrained scenarios. To address these limitations, we propose a lightweight dual-domain multi-scale feature network (DMFN) for high performance blind image watermarking. Our model adopts a symmetric encoder-decoder architecture with an overall UNet-like structure, enhanced by depthwise separable convolutions and dilated convolutions to better preserve watermark imperceptibility while improving robustness against various noises. To efficiently capture global features without relying on computationally expensive self-attention modules, we introduce a dual-domain feature block (DDFB), which enables global feature learning with reduced model complexity. Additionally, we incorporate an FFT-domain loss during the training process to further improve the imperceptibility of watermarked images, and adopt a multi-scale feature extraction architecture to better capture both image and watermark information across multiple levels. Extensive experiments demonstrate that our method outperforms state-of-the-art watermarking schemes in terms of both robustness and imperceptibility, while maintaining a compact model size and efficient inference speed.