Conditional Score Matching for Image to Image Regression
Hao Xin, Michael Yu Zhu · 2020
Image to image regression is an essential machine learning problem that is commonly encountered in computer vision tasks. Previous research works have been concentrating on task-dependent end-to-end regression models. Although such a discriminative approach has shown its proficiency in producing precise predictions, its architecture is usually overly complicated for interpretation and unable to provide uncertainty quantification for prediction. We propose a general image to image regression framework named CSMNet which is based on conditional score matching and Langevin dynamics. The proposed framework is generative and emphasize on the probabilistic perspective of general image to image regression. Discriminative models are also involved as sufficient feature extractors to complement the generative modeling of the distribution of the target image conditioned an input image. The proposed model is trained by multi-level denoising score matching. Subsequently, the learned model is used for sampling from the conditional distribution via annealed Langevin dynamics. We introduce a grid selection procedure of Langevin sampling which can produce robust prediction. Comprehensive experiments are conducted on diverse image to image regression problems. The proposed model surpasses other purely generative models and achieves comparable performance to some end-to-end regression methods. Our method also demonstrates remarkable generalization ability which further demonstrates the advantage of the proposed conditional score matching framework.