Abstract: The images of typical open-pit mining scenarios exhibit multi-type composite noise characteristics, with a low signal-to-noise ratio and significant spatial heterogeneity. Most existing deep learning models directly transfer denoising architectures from natural images, ignoring the unique noise distribution patterns of mining remote sensing images. To address the issue, a mine remote sensing image denoising method based on improved YOLOv5 was proposed. Considering the instability of traditional YOLOv5 in high-noise environments, a multi-scale feature fusion module was introduced to enhance the model's ability to recognize noise of different sizes. Additionally, a residual attention mechanism was incorporated to improve the extraction of useful features and enhance the robustness of the denoising effect. An adaptive noise estimation technique was employed to dynamically adjust denoising parameters based on the noise characteristics of different image regions, achieving more precise noise suppression. The experimental results showed that the improved YOLOv5 significantly outperformed other classical denoising methods in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). Compared to the original YOLOv5, the PSNR value increased by 2.5 dB, and the SSIM improved by 0.05. The improved YOLOv5 performed well under all noise types, especially in Gaussian noise environments, where its PSNR and SSIM reached 32.5 dB and 0.95, respectively, significantly surpassing other classical denoising methods.