Test drives for the development of camera-based automotivealgorithms like object detection or instance segmentation are veryexpensive and time-consuming. Therefore, the re-use of existingdatabases like COCO or Berkeley Deep Drive by intentionallyvarying the image quality in a post-processing step promises tosave time and money, while giving access to novel image qualityproperties. One possible variation we investigate is the sharpnessof the camera system, by applying spatially varying optical blurmodels as low-pass filters on the image data. Any such opera-tion significantly changes the amount and distribution of noise, acentral property of image quality, which in this context is an un-desired side-effect. In this article, a novel method is presentedto reconstruct the original camera sensor noise for the filteredimage. This is different from denoising. The method estimatesthe original camera sensor noise using the combination of princi-pal component analysis (PCA) and a variance-stabilizing trans-formation. The noise is then reconstructed for the filtered imagewith the PCA applied locally on small image sections, and an in-verse variance-stabilizing transformation. Although the resultingnoise distribution can slightly deviate from the original, this novelmethod does not introduce any image artifacts as denoising woulddo. We present the method as applied to synthetic and real driv-ing scenes at different noise levels and discuss the accuracy of thereconstruction visually and with statistical parameters.