Depth of field post-processing using neural networks

Sammanfattning

It is possible to create images from a computer model closely resembling those taken with a physical camera. To improve the photo-realism and perceived quality of a rendered image, it is often desireable to add realistic effects that do not appear in computer graphics due to the camera model. One of these effects is depth of field, where a physical camera with a lens can only focus at one specific distance in a scene, making the rest of the scene appear blurry. This is opposed to a rendered image, where the entire scene appears sharp. In this thesis work, it is investigated if a neural network is able to replicate the depth of field effect in computer rendered images, when given an image and the distance from the camera of objects in the image. Three neural networks based on the same structure are created and studied, and it is found that the simplest model fails to retain background information in the images while the most complex model manages to replicate the depth of field effect, with an average PSNR of 43.12 and average SSIM of 0.987.