Image Recognition Based On Convolution Neural Network And Its Implementation In Spark Platform

In recent years,the relevant research works of deep learning developed rapidly.It has a wide range of potential applications in various domain.Especially in image recognition,it has achieved a great progress.With the increase in image size and the complexity of the neural network structure,the stand-alone CPU computing hit a bottleneck.On the other hand,the use of multi-core GPU architecture can reduce the training time and improve the accuracy,which is suitable for large-scale multi-task operation.However,as the size of data and the complexity of network structure increased further,training neural network in a stand-alone GPU cannot meet the application requirements under the circumstance of big data.To address the problems,this thesis tries to improve the structure of existing convolutional neural network,and design a distributed deep learning architecture in Spark platform.It is hoped that the proposed model and architecture can improve the training efficiency and recognition accuracy of the existing neural network models.Based on this,the main content of this thesis include:(1)This thesis analysis the training algorithm of convolutional neural network(CNN).By comparing the merit and disadvantage of different type of models,the thesis proposed an improved CNN model.(2)This thesis analysis the existing distributed deep learning architecture,and proposed two data-based parallel schema:DOWNPOUR and ADAG parallel algorithm.(3)Based on the proposed model and algorithm,the comparative experiments are conducted on CIFAR and MNIST dataset to provide an empirical analysis.By experimental analysis,the thesis compares the performance of the distributed ADAG and DOWNPOUR algorithm with different running parameters.The comparative results have proven that,when the number of worker node in distributed platform increases,the required training time of neural networks will reduce.In addition,the proposed CNN model will also improve the accuracy of image recognition in distributed platform.