Design And Application Research Of Convolutional Neural Network Accelerator Based On ZYNQ Platform

In recent years,deep learning has emerged as a new form of machine learning that enables computers to learn from experience and understand the world in terms of conceptual levels.As a new artificial neural network method,the convolutional neural network(CNN)makes it possible to share the weights between neurons to reduce the training parameters of the samples so that the generalization ability and accuracy of the classification can be further improved.Therefore,CNN has been widely promoted and applied successfully in the field of image recognition.At present,the main way to realize CNN is usually based on general-purpose processors,but this software-based approach does not allow CNN's parallelism to be fully exploited and enables applications to have real-time,flexibility and power consumption needs can not be satisfied.In addition,since any CNN model cannot be optimally generalized for all datasets,a suitable set of hyperparameters must be selected before applying CNN to a new dataset.Choosing a new model for a new data set can be a time-consuming and tedious task.The number of adjusted hyperparameters and the evaluation time of each new hyperparameter set make their optimization in the CNN model particularly difficult.The project adopts a SOC product ZYNQ platform that is supported by software and hardware co-design by Xilinx.The development of ZYNQ will benefit both us ARM's rich ecosystem resources,in turn,can benefit from the flexibility and scalability of FPGAs.This thesis introduces the development mode and convolution neural network of ZYNQ platform in detail,designs CNN hardware accelerator,and proposes a CNN hyperparameter optimization method based on improved Bayesian optimization algorithm.This method uses the improved Thompson sampling method as the acquisition function,and uses the improved Markov chain Monte Carlo algorithm to accelerate the training of Gaussian agent model.The method can perform hyperparameter optimization under different CNN frameworks with hyperparameter space.Using CIFAR-10,MRBI and SVHN test set to test the performance of the algorithm,the experimental results show that the improved CNN hyperparametric optimization algorithm has better performance than the similar hyperparametric optimization algorithm.Finally,the CNN hardware accelerator and CNN hyperparameter optimization algorithm are integrated into the software and hardware collaborative system for experimental verification.Compared with the PC-side convolutional neural network,the convolution neural network based on the ZYNQ platform has better performance.Thus displaying the advantage of the image recognition technology based on the ZYNQ platform convolution neural network in the aspect of parallelism and more broad application prospects.Therefore,it is of great academic value and application value to carry on the research of this topic.