Research On Image Sampling Based On Reinforcement Learning

With the advent of digital information age and the development of multimedia technology,the amount of image data used by people is increasing day by day.The research of image representation has been paid much attention.The image representation method is the image representation and storage way on the computer.The effective use of some pixel points of the image can save the storage space needed for an image representation,and improves the speed of image processing.Image sampling technology can collect pixel information,so image sampling technology is provided to extract a lot of image data.The traditional triangular mesh sampling method sparse samples the image by generating triangular mesh so as to obtain the optimal sampling point set.However,its disadvantage is that the triangular mesh generated by the sampling points will form a convex hull,and some key contents of the image can’t be accurately obtained,thus affecting the quality of image restoration.The random walk sampling method can focus on local information.However,it also has an obvious drawback that the random walk sampling method can’t ensure the quality of image restoration and is prone to over-sampling or under-sampling.Therefore,an image sampling algorithm based on reinforcement learning is proposed in this thesis.The agent interacts with the environment and uses the experience knowledge to train for updating the strategy so as to obtain the optimal strategy.Experimental results show that the algorithm proposed in this thesis can not only sample the image effectively,but also ensure the quality of image restoration.The main research contents of this thesis include the following two points:(1)On the basis of research on the existing image sampling methods,aiming at the problems existing in the traditional image sampling methods,this thesis combines image sampling with reinforcement learning,and proposes a new image sampling algorithm.Reinforcement learning agent train and search sample subset by "trial and error".Then agent use image restoration model to evaluate the sample subset and use the feedback to adjust the sampling strategy of the agent,finally agent gets the optimal strategy and the optimal image sampling results.Experiments show that the algorithm can not only sample the image effectively,but also ensure the quality of image restoration.(2)Through comparative experiments,the feasibility and effectiveness of the image sampling algorithm based on reinforcement learning proposed in this thesis are verified.However,compared with the traditional image sampling method,there is a problem of too few samples in areas with rich image textures,which have a certain impact on the quality of image restoration.Aiming at the problem,the algorithm in this thesis is enhanced by the gradient operator and the farthest point sampling algorithm.Experiments show that the improved algorithm has better image restoration quality than the original algorithm and the traditional algorithm.