| Deep Learning(DL)is a growing topic of interest over the last decade due to its propagation to people’s daily lives.While cloud computing has contributed to the provision of a wide range of services to the end users/devices,a rising expectations are observed from the service petitioner in terms of minimizing the service transmission latency to obtain the superb service experience at way less cost than traditional on premises systems.Cloud computing is applied to facilitate service range between the device and the internet,in which the cloud nodes are placed in the proximity of the devices to provide different resources needed for application deployment and service operations.Furthermore,the autonomous data collection and processing capability at cloud has attracted attention from many data-driven deep learning applications.It starts to draw widespread attention owing to the rapid development of smart solutions,all of which benefit from the advances of deep learning technologies.Specially,different deep learning models are designed in accordance with the service requirements initiated at the cloud,such as real-time character detection,healthcare monitoring,and face recognition,etc.While most of the current works focus on the improvement of the model structure and training optimization on the centralized cloud server,deep learning aims to scatter the learning and inference tasks mostly in the cloud environment to expedite both the data processing and model training so that the intelligent service could be rendered in a timely manner.It is thus desirable to explore the collective behavior of cloud computing to provide an all-round enhancement for the cloud based deep learning.By doing so,the deployed nodes at the edge help complement the inadequacy of the cloud through collaborations to full complex tasks while presenting flexible scalability and enhanced robustness in the deep learning context.In this thesis,we are committed to investigating how to use cloud computing to realize the intelligent service provision.The whole research work could be considered as development of cloud-based deep learning multi-service platform for segmentation and detection on various data sets to meet the rigorous quality of service requirement in the network training and deliver the real-time intelligent services under the designated platform which composed of several phases.We started by focusing on building the cloud-based platform for handwritten character recognition using MNIST dataset to support heterogeneous deep learning applications in the first phase.The second phase concentrates on sentimental analysis on IMDB movie review dataset by Stanford using state of the art cloud based deep learning model.The next phase concerns with cloud based facial recognition on android devices,aiming to enhance the data processing efficiency at the edge and facilitate the learning-based services.Finally,the last phase is to design and generalize the training process of 3D U-Net deep learning model for semantic segmentation of Multimodal multisite MRI dataset through the collaborations of cloud to achieve a well-performing model without compromising the privacy.The thesis advances the state-of-the-art by making the following contributions:1.A novel cloud-based deep learning DL4 J model is designed,our suggested architecture for hand written digits dataset having training set of 60000,and a test set of 10000 images shows high performance in terms of accuracy and time.Hence,in our proposed system,handwritten characters are identified with 99.41% accuracy with less computation time.2.For the sentimental analysis of movie reviews a binary dataset consisting of50,000 reviews from Internet Movie Database(IMDB)is used to propose a state-ofthe-art cloud based deep learning model consisting artificial neural network(ANN)having 6 layers(4 hidden,1 input and 1 output).The training accuracy of this model has reached 91.9% and validation accuracy which is 86.67%.3.The information-centric cloud based facial recognition system is proposed,designed and integrated,from which real time data communication is enabled for efficient data processing at internet cloud.4.We presented a unique cloud-based 3D U-Net method to perform brain tumor segmentation using Multi-modal multi-site MRI(FLAIR,T1 w,T1gd,T2w)dataset consisting 750 volumes.The system was effectively trained by utilizing multiple hyper parameters.We got an average dice score of 95% which makes our method the first cloud-based method to achieve maximum accuracy.”... |
PDF Full Text Request