Traditional deep learning technologies often struggle to balance local details and global structural features in the super-resolution reconstruction of structure illumination microscopy (SIM) images, with limited performance particularly in sparse input and low illumination conditions. To this end, this study conducts wide-field microscopic imaging experiments under varying illumination intensities and constructs microscopic image datasets in different lighting conditions. A hierarchical-fusion feature aggregation network based on multi-scale convolution (MSCF-HFNet) is proposed. By leveraging an encoder-decoder structure integrated with multi-scale feature extraction and fusion mechanisms, MSCF-HFNet achieves efficient mapping from low-frequency to high-frequency images, significantly enhancing reconstruction capability under low-contrast and complex structural scenarios. Experimental results demonstrate that MSCF-HFNet outperforms competing models (EDSR, scU-Net, and DFCAN) with average improvements of 5.81% and 4.60% in PSNR and SSIM respectively, and a 16.81% reduction in MSE. Additionally, FWHM-based resolution evaluation indicates that MSCF-HFNet achieves spatial resolution below 120 nm, featuring an 13.89% improvement compared to other models. Simultaneous dataset simulation confirms that the model achieves a sound balance between model complexity and reconstruction accuracy across datasets, demonstrating excellent robustness and consistency.