The quality of three-dimensional (3D) reconstruction algorithm Structure from Motion (SfM) is affected by the input image’s resolution and CMOS’s noise level. We propose a denoisable Super Resolution (SR) method to improve resolutions while denoising for SfM’s input images taken by a CMOS device, improving its performance on noisy images. The conventional deep learning SR algorithm does not consider denoising during the learning process. This results in the disability of simultaneously reducing noise and improving resolution. In our methods Add Noise before Downsampling (An-Ds) and Downsampling before Adding Noise (Ds-An), instead of expanding the training data, we extract the noise from a real-world noise dataset and selectively add it to low resolution (LR) images of the SR training set. Thus the SR algorithm can simultaneously improve resolution and reduce noise. Moreover, selectively adding noise also remain SR’s performance on clean images. We trained two representative SR algorithms (SRCNN and EDSR) using traditional and our designed methods to process both clean and noisy images. Without changing the SR network’s structure, improvements of 0.17 dB in Peak Signal Noise Ratio (PSNR) by Ds-An and 0.14 dB by An-Ds (approximately 20% of improvement in three years) were observed in noisy images’ experiments by EDSR. Meantime, there is only a little loss (less than 0.01 dB) on the clean images. The SfM’s results show a better reconstructed 3D model using our methods. Compared to non-preprocessing and conventional preprocessing, key metrics such as Mean Reprojection Error (MRE) reduced 51.9% and 12.6%, and 2D key-points matching rate improved 41.7% and 217%, respectively.