Deep Unrolling Based Image Layer Separation And Application

Images captured in natural scenery often suffer from bad weather condition,such as: rain,haze,snow and light,resulting in the degraded outdoor images.Existing methods for solving these problems can be divided into two categories: conventional model optimization based methods and deep learning based methods.Conventional optimization methods rely on complex prior constraints designed on specific task,resulting in the huge time consumption.Moreover,complex prior constraints require finely mathematical abstraction of specific problems which requires extremely high mathematical skills and a deep understanding of the problem.Deep learning based methods rely on the black box network model established on training data,such methods are not suitable for complex scenarios and other tasks,and their generalization ability is weak.We find that lots of tasks can be abstractly understood as the separation of two independent components.At present,the solutions to these problems are based on specific modeling of the respective problems,and they cannot be transferred to other tasks.In this paper,we propose a deep unrolling introduced layer separation framework for these problems,in which we introduce an adaptive feature learning method to meet the needs of multi-task and own theoretical guarantees.At present,the algorithm has been applied to four applications: single image rain streak removal,single image reflection removal,low-light image enhancement,and tone mapping.Experimental results verify the effectiveness and generation of our algorithm.Inspired by layer separation,we also propose a method to solve task-aware image deraining and image dehazing by combining traditional model priors and deep neural networks.For image rain streaks removal,this method not only retains the ability to describe natural scenes with total variation model constraints,but also implements the learning of complex interference features based on a learnable residual network.For image dehazing,we integrate the data-driven and domain knowledge to solve the estimation of transmission,which is the core point of image dehazing.Besides,focus on the phenomenon of rain and haze coexisting in heavy rain environment,we extends the atmospheric light reflection model to make it more suitable for the real rain environment.Comparing with the existing methods,it can be found that our method can solve the heavy rain situation where rain and haze are both existing,remove the rain streak interference to a great extent,retain rich background details and restore the real scene color,verifying the effectiveness and practical value of the this paper.