ProxImaL: Efficient Image Optimization using Proximal Algorithms

摘要

Computational photography systems are becoming increasinglyrndiverse, while computational resources—for example on mobilernplatforms—are rapidly increasing. As diverse as these camera systemsrnmay be, slightly different variants of the underlying imagernprocessing tasks, such as demosaicking, deconvolution, denoising,rninpainting, image fusion, and alignment, are shared between all ofrnthese systems. Formal optimization methods have recently beenrndemonstrated to achieve state-of-the-art quality for many of thesernapplications. Unfortunately, different combinations of natural imagernpriors and optimization algorithms may be optimal for different problems,rnand implementing and testing each combination is currentlyrna time-consuming and error-prone process. ProxImaL is a domainspecificrnlanguage and compiler for image optimization problemsrnthat makes it easy to experiment with different problem formulationsrnand algorithm choices. The language uses proximal operators asrnthe fundamental building blocks of a variety of linear and nonlinearrnimage formation models and cost functions, advanced image priors,rnand noise models. The compiler intelligently chooses the bestrnway to translate a problem formulation and choice of optimizationrnalgorithm into an efficient solver implementation. In applicationsrnto the image processing pipeline, deconvolution in the presence ofrnPoisson-distributed shot noise, and burst denoising, we show thatrna few lines of ProxImaL code can generate highly efficient solversrnthat achieve state-of-the-art results. We also show applications tornthe nonlinear and nonconvex problem of phase retrieval.