A Novel Perceptual Metric in Deep Learning

Sammanfattning: Loss functions are a crucial part of image processing when using modern neural networks, trained with stochastic gradient descent. There exist multiple loss functions today, some claiming to be perceptual. Researchers at NVIDIA recently published a proposal of such a metric called FLIP. This led to our work on this thesis where we present a comparison between multiple loss functions, both established ones but also a completely new one. The loss functions were subjected to two problems, denoising RGB images and reconstructing MR images. The central question is how does one evaluate the evaluation? If we train a network with the same architecture but with different loss functions, and then measure the performance with one of these loss functions, the result would most likely be biased. We therefore present a comparison between loss functions where we both visually and numerically evaluate the results as well as conducting user studies. We find that the weighted version of LPIPS + l2 and MS-SSIM are especially good loss functions for mentioned problems, and the newly proposed metric FLIP performed well.