Research On Low-level Vision Image Processing Method Based On Multi-task Deep Learning

The low-level visuals such as image super-resolution,de-blurring,re-colorization,etc.are the basis of high-level visual tasks.In recent years,due to the outstanding performance of deep learning technology in high-level visual tasks,the job of applying it to the low-level visual tasks has gradually emerged.However,these current deep learning low-level image processing methods are usually only suitable for a single task or to a single factor.Faced with multiple low-level visual mission requirements,such as the need to consider multiple image degradations simultaneously — resolution reduction and motion blur,these single-tasks work less well.Based on this,this paper will utilize the multi-task deep learning method to try to jointly deal with a variety of low-level visual tasks and explore the corresponding learning models and application modes.This paper first introduces the principles of deep learning and the low-level visual image processing tasks involved,and then analyzes the low-level visual depth learning methods of related frontiers.Finally,different depth learning models are proposed to achieve multi-task joint image processing.(1)Single satellite imagery simultaneous super-resolution and re-coloring using multi-task deep neural networksDue to the limited capabilities of hardware devices or ground camouflage,most satellite images have problems such as low spatial resolution and monotonous color.To this end,a multi-tasking deep network was constructed to simultaneously implement two different image tasks,super-resolution and re-colorization.The experimental results verify the effectiveness of the multi-task deep learning model.(2)Cooperative Learning a Deep Network for Single Image Simultaneous Super-resolution and Motion DeblurringImage quality degradation may be caused by factors such as reduced spatial resolution and motion blur.For this reason,the multi-factor image degradation process is analyzed,and a deep network model collaboratively learned by two sub-networks is constructed based on the degradation analysis to jointly solvemulti-task problem.The collaborative learning deep model has been experimentally verified to have excellent image recovery performance.(3)Cooperative training model based on satellite image super resolutionThe cooperative training network overcomes the problems of training instability and mode collapse in Generative Adversarial Networks(GANs)by alternately training the generator and the mediator.In view of this,the cooperative training network was improved and applied on the satellite image super-resolution task.The experimental results show that the improved cooperative training network can work effectively.