Research On Image Denoising Convolutional Neural Network Based On Evolutionary Algorithm

As a preprocessing procedure of medical images,medical images denoining is of great significance for subsequent medical diagnosis and treatment.As a widely used neural network,convolutional neural networks(CNN)has excellent image denoising capabilities.Where in the setting of CNN hyperparameters and structure is essentially an optimization problem.Therefore,this thesis focuses on evolutionary algorithms and applies them to the optimization of hyperparameters and structures of image denoising CNN.The main contents of this thesis include the following aspects:Considering the problem that the hyperparameters and structures of CNNs for image denoising are difficult to determine,a CNN hyperparameter and structure optimization method based on teaching learning based optimization(TLBO)algorithm and differential evolution(DE)algorithm is proposed.In the early stage of the evolution process of the hybrid TLBO-DE algorithm,the population evolved with a high probability using the evolutionary mechanism of the DE algorithm,thereby improving the population diversity.In the later stage of evolution,the population evolved with a high probability using the “teaching” mechanism of the TLBO algorithm,thereby improving the convergence speed of the algorithm.Finally,the proposed method is tested on the CEC2017 test suite and the public medical image dataset.Experimental results show that compared with the state of the art algorithms,the proposed hybrid TLBO-DE algorithm has higher convergence speed and stability.Compared with CNN denoising methods based on genetic algorithm(GA),DE and TLBO,the proposed method has better optimization performance and image denoising performance.Compared with the current block matching and 3D filtering and denoising convolutional neural networkmethods with good denoising performance,the proposed method has better denoising performance.A hybrid evolutionary algorithm is proposed,which is applied to the optimization of hyperparameters and structures of denoising CNN.The hybrid algorithm combines the marine predators algorithm(MPA)、the whale optimization algorithm(WOA)、the white shark optimizer(WSO)with the “teaching” stage of the TLBO algorithm.In the evolutionary process,it adopts four subpopulations with different sizes for co-evolution,which effectively improves the global search ability of the algorithm.In addition,in the evolutionary process,the algorithm determines the evolutionary mechanism adopted by each subpopulation according to the comprehensive optimization performance of the four evolutionary mechanisms,which improves the convergence speed and convergence accuracy of the algorithm.The proposed method is tested on the CEC2017 test suite and the public medical image dataset.Experimental results show that compared with the state-of-the-art algorithms,the proposed hybrid evolutionary algorithm has higher convergence speed and stability.Compared with CNN denoising methods based on particle swarm optimization(PSO)、MPA、WOA and WSO,the proposed method has better optimization performance and image denoising performance.A multi-objective hybrid evolution algorithm is designed for the multi-objective optimization of CNN network structure and hyperparameters.It introduces nondominated sorting,crowding distance calculation and elite strategy into the designed hybrid evolutionary algorithm to realize the multiobjective optimization of image denoising criteria.Where in the nondominated sorting improves the convergence speed and robustness of the algorithm,and ensures the uniform distribution of the pareto optimal solutions.The crowding distance calculation preserves the diversity of pareto optimal solutions.Meanwhile,the elite strategy uses external archive population to preserve individuals with better fitness values,which improves the computational efficiency of the population.The proposed method is tested on a public medical image dataset.Experimental results show that compared with CNN denoising methods based on multi-objective DE,GA,PSO,WOA and WSO algorithms,the proposed method can give a better pareto approximate optimal solution set and has better image denoising performance.