Reasearch Of Deep Optimization Networks-based Image Enhancement Algorithms

In this article,we discussed how to design the deep model which can solve the image enhancement problem from the optimization perspective.The models we proposed are incorporated with the knowledge gained from conventional model and show their high performance and interpretability in various experiments.With the significant success of Convnets in the high level computer vision community,some researchers also introduce it into the low level vision community.However those deep Convnets often designed by various heuristic method and empirical intuition,unlike those conventional methods which can incorporate various domain knowledge.We want to design some deep models with the optimization problem succeed in conventional method as the guidance.We proposed two kind of deep models which combined with popular optimization techniques and image domain knowledge.Those models are both applied to real world image restoration tasks.The first model is the deep hybrid residual network with stastical prior which proposed to solve image super-resolution problem.We design this deep model by the stimulation of gradient-based optimization method which solves the MAP estimation of original degradation model.The second method is proposed to enhance the recent existing deep models which created by unroll the optimization of degradation model.These deep models are popular because they use powerful Convnets to replace the one step in conventional optimization method.Although they benefit from these Convnets,they also lost the well property of the original optimization techniques because of intuitive replacement.We improved such models so that they can have the theoretical convergence property again.In experiment part,we will verify the model which can be shown not only they have good convergence property,but also have the good results on image deconvolution problem.