Image Reconstruction Algorithm Based On Simplified Numerical Optimization Back Propagation Neural Network

With the continuous progress of the times,people have much higher requirements on image,quick access to high-definition image has become a basic task in the field of image information research.Therefore,the requirement for the reconstruction of optical CT image is also increasing.In recent years,the application of BP neural network algorithm in optical CT image reconstruction to quickly acquire high precision image has become a research hotspot.BP neural network is specifically a kind of neural network which is trained according to the error back propagation,it is based on multi-layer feed forward.Because of some deficiencies of itself,such as the convergence rate of the network is slow,the training process is complex and unstable,easy to fall into local minima and so on.Aiming at the above problems of BP neural network,the domestic and foreign research scholars have improved the BP neural network algorithm,the improvement method is mainly divided into two categories: one improvement of the method is based on the gradient descent method,and the other method is based on the numerical optimization method.The improvement of the gradient descent method mainly includes the change of transfer function,the additional momentum factor and the adaptive learning method.Among them,the numerical optimization method to improve the accuracy and convergence rate compared to the standard BP algorithm has been greatly improved,however,the improved method can improve the accuracy of the same time,the calculation of the required memory is very large,and the iteration process is complex,the small and medium-sized computer can't complete the reconstruction calculation.In this paper,the problem of the numerical optimization method is simplified,and the conjugate gradient method,the quasi Newton method and the LM algorithm are simplified,In the case of improving the convergence speed and reducing the memory consumption,the error accuracy error of the calculation is guaranteed within the acceptable range of image reconstruction.Specific work is as follows:(1)The search factor ?)(k in the conjugate gradient method is simplified by a constant value of 1,in the process of iterative calculation,it is required to repeat the operation of the adjacent two times in the direction of the gradient,which makes the memory consumption is large,and the speed is slow,the simplified conjugate gradient method has a faster convergence speed,the deviation of the calculated error value is only 10-5 order of magnitude compared to that of the not simplified conjugate gradient method.In thequasi Newton method,the mean method is simplified,and the deviation of the calculated error value is also obtained for the 10-5 order of magnitude,the memory consumption of the quasi Newton method is reduced by nearly half.The iterative weight parameters of LM algorithm are simplified,and the simplified LM algorithm not only shows the advantages of the quasi Newton method with high precision,but also obtains the advantage of the conjugate gradient method.(2)Studying the three kinds of simplification algorithms applicability,the simplified quasi Newton method is suitable for the medium scale neural network structure,the simplified quasi Newton method is suitable for the medium scale neural network structure,the simplified conjugate gradient method is suitable for large scale neural network structure.The application of the simplified LM algorithm in small scale image reconstruction is mainly studied,the small scale Shepp-Logan model is selected as the analysis object of image reconstruction,shows in the small scale structure of the neural network,the LM algorithm is simplified in the case of error(10-6),and its convergence rate is obviously improved.