Self-supervised image denoising has received extensive attention because of its ability to train only noisy images. Recently, Asymmetric PD and Blind-Spot Network (AP-BSN) has been proposed to eliminate the spatial correlation of noise. However, when the noise connection area is large, using the single central pixel mask will cause the recovered blind-spot pixels still contain noise, resulting in obviously abnormal color spots in the denoised image. In addition, simply stacking dilated convolutional layers introduces block artifacts, which can destroy the high-frequency details of the image. To address the above issues, we propose an asymmetric mask convolution kernel (AMCK), forming a large diagonal blind-area during training to further decouple the spatial connection of large-scale noise, and a small blind-spot during the inference process to minimize detail information loss. We also propose a detail information supplement (DIS) module to minimize damage to the image structure and the high-frequency information. Extensive quantitative and qualitative evaluations of the SIDD and DND datasets demonstrate that our method comprehensively outperforms other self-supervised models in denoising and image texture maintenance.