<p>Research on image watermarking algorithms mainly focuses on enhancing the capacity of embedded information, imperceptibility, and the ability to resist various attacks and interference. Currently, a common method to effectively enhance the robustness of watermarks is to insert an attack simulation layer after the watermark image. However, this method is not effective against undifferentiable noise such as JPEG compression. In this paper, we propose a robust watermarking algorithm based on dual-branch and adaptive noise weighting, which can achieve an end-to-end watermark network training process. By combining cover-dependent embedding and non-cover embedding methods to improve the message embedding process, and using an adaptive weighting noise layer to improve the training method, the network can learn more efficient embedding methods and more robust message encoding simultaneously. In addition, a distortion reconstruction module is added to the extraction network to recover the information loss caused by noise, further enhancing the extraction effect of the watermark information. The method proposed in this paper can achieve an extraction accuracy of 99.99% with an embedding length of 30 bits and a PSNR of 45.36dB. Especially in the face of undifferentiable noise situations such as JPEG compression, the robustness of the watermark is significantly improved.</p>