In seismic exploration, field data are always affected by noise, reducing the quality of subsequent analysis. Striking a balance between signal preservation and noise attenuation effectively is a major challenge. This study adopts a dual-branch framework with convolutional neural networks (CNN) and Transformer architecture that integrates blind-spot convolution (BS-Conv) for local detail capture, rank-augmented linear attention (RALA) for global feature extraction, and the gated CNN block (GCB) for adaptive feature selection to achieve high-fidelity signal reconstruction and noise suppression. Specifically, the CNN branch introduces BS-Conv to strengthen the local feature extraction of seismic data. The Transformer branch adopts RALA to capture long‑range dependencies, strengthening global feature extraction. Finally, each skip connection is equipped with GCB to adaptively extract relevant features. Both branches are constructed on the U-Net++ architecture and utilize its multi-scale feature-fusion capability to optimize denoising performance. The network was tested on synthetic data and field seismic data to validate the model performance. We quantitatively evaluated the denoising results using peak signal‑to‑noise ratio and structural similarity index. Comparative tests prove that this dual-branch denoising framework is superior to other methods in both noise reduction and signal fidelity. These results further confirm its effectiveness and robustness for seismic denoising on both synthetic and field data.