<p>Building on recent research, supervised learning methods excel in synthetic image denoising but often fall short on real-world noisy images due to reliance on noisy-clean image pairs and synthetic-real noise discrepancies. In this paper, we propose a novel self-supervised method, named Regularized Multi-Pattern Blind Spot Network (RMP-BSN). The key innovation of our method lies in the Multi-Pattern Fusion (MPF) module with the Spatial Group-wise Enhance (SGE) mechanism, which enables simultaneous use of multiple convolutional kernel shapes. This approach effectively removes noise correlation, restores texture structures, and enhances information transmission. Additionally, we design a regularized loss function to mitigate overfitting and improve denoising. Comprehensive experiments on real-world datasets underline the superiority of the proposed RMP-BSN over the state-of-the-art self-supervised real-world image denoising methods.</p>