High-capacity steganographic schemes often involve considerable modifications to the cover image structure, which can adversely affect its visual fidelity and undermine the imperceptibility requirement. Conversely, prioritizing visual fidelity typically restricts the embedding capacity. To address this trade-off, this paper proposes a color image steganography scheme based on a GAM-CBAM dual-attention mechanism. The proposed method integrates two key attention modules: the global attention module (GAM) enhances feature modelling capability, whereas the convolutional block attention module (CBAM) enables precise embedding region localization and dynamic feature weight assignment, effectively suppressing redundant information and reducing detection risk. The encoding network adopts an improved U-Net architecture with global attention to optimize feature extraction, whereas the decoding network incorporates CBAM to increase reconstruction accuracy. Experiments conducted on the COCO dataset for 256 × 256 color image steganography demonstrate the scheme’s superiority: PSNR improvement of 2.1687 dB (cover/stego image) and 1.489 dB gain (secret/reconstructed image) at a 1-byte per-pixel embedding capacity. These results validate the method’s significant improvements in visual fidelity and information recovery accuracy compared with the prior U-Net-based methods.