Abstract: Previous deep learning-based denoising methods struggled to balance high-quality image details and color features in complex denoising scenes such as high-frequency detail regions and weak-texture regions. To address this issue, a multi-scale feature fusion denoising network based on reduced biquaternion (RQ) is proposed. The network utilizes RQ convolutional and deconvolutional layers as primary feature extraction blocks to capture and restore image features, which can effectively obtain inter-relationship between channels and learn color prior information of images. Additionally, a reduced biquaternion channel attention block (RQCAB) is designed to fuse multi-scale feature information, so as to fully extract local detail features of images. By introducing a new hybrid loss function to supervise the training process, the network achieves better visual results. The denoising experiments on three different denoising datasets demonstrate that the proposed method achieves the highest average PSNR of 31.08 dB, surpassing the previous state-of-the-art method KBNet by 0.7 dB, with a reduction in parameter and computational cost to 2.49 × 10⁶ and 25.00 × 10⁹, making it a lightweight and efficient denoising approach.