The BP Neural Network Image Reconstruction Algorithm Based On Prior Knowledge

As the application of CT reconstruction technology more and more widely,the requirement of reconstruction is rising at its speed and precision,so the algorithm requires to improve.Especially for the complicated situation,although the combination of improved reconstruction algorithm has high reconstruction accuracy,but it requires repeated iteration and cannot monitor the dynamic field in real-time.Spectrum image reconstruction has a high data processing speed,but the reconstruction unable to meet the needs of the real-time measurement because of low precision and poor stability.Neural network is more suitable for this kind of image reconstruction of complex field,intelligent,imitation of biochemical,parallel processing and information fusion is the trend of tomography image reconstruction algorithm.Artificial neural network is made up of many functions of parallel processing unit,because of its special structure and strong function of information process,so people is hot on the study of artificial neural network.There are many model of artificial neural network,BP neural network is one of the most widely used model,it is a kind of classical neural network model,its structure is simple and can approximate arbitrary function in theory.Under the condition of the BP neural network is used in a few projection of CT image reconstruction,not only can reduce the time of the image reconstruction,but also can improve the accuracy of image reconstruction.However,there are several obvious shortcomings in classical BP neural network model: convergence time is too long,easy to fall into local minimum,the poor generalization ability of the network.For the defects of BP neural network,this article from the following several aspects to study the BP neural network based on prior knowledge:(1)Analysising process of topological structure and algorithm of BP neural network,point out the inherent defect of the BP neural network,on this basis,analyzed with the momentum method and adaptive learning rate method,elasticity of BP algorithm,quasi-newton method,conjugate gradient method,the LM method such as the current several improved methods.(2)Put forward a new method of BP neural network to build the BP neural network algorithm based on prior knowledge.The specific content of this method is to apply prior knowledge to the selection of BP neural network model,the calculation of excitation function,the network weights initialization and network learning on the sample.(3)Building a kind of BP neural network model based on prior knowledge and apply to the CT image reconstruction.The model will be the projection values of the original image as input data to the network,the reconstruction of the image projection values as the output of the network data,the desire of reconstruction image for the hidden layer of network data.The model from image reconstruction theory,using the projection matrix as neural network weights between hidden layer and output layer of the matrix,in the training of network only adjust the weights between input layer and hidden layer,which reduces computation to speed up the network convergence speed.In this paper,based on the prior knowledge of the BP neural network model is apply to the simulated image reconstruction on the Shepp-Logan model,the image reconstruction results compared with the result of traditional image reconstruction algorithms available: based on the prior knowledge of the BP neural network algorithm can greatly reduce the network convergence time,and improve the accuracy of image reconstruction.