Over the past years, deep neural networks have proved to be an essential element for developing intelligent solutions. They have achieved remarkable performances at a cost of a large size with deeper layers and millions of parameters. Therefore utilising these networks for developing augmented reality applications on limited resource platforms such as embedded devices or mobile phones is a challenging task. In this context, this thesis addresses the problem of neural networks compression and optimisation in order to enhance the performance of these models on limited resource systems.In the first part of this manuscript, we present an overview of different compression methods present in the literature as well as their strengths, their weaknesses and a brief comparison of these techniques. We are also interested in the methods that are allowing the optimisation of deep neural networks structure design, from simple modules to autonomous models building.In the second part, we show the feasibility of an augmented reality application in real-time using a deep learning model in order to achieve a face parsing task. By using adapted frameworks and an optimised architecture, we achieve the segmentation of different face components in real-time with a high level of consistency on an iPhone X.The last two parts are focused on the development and the evaluation of a new deep convolutional neural networks compression method. Based on a regularisation term which is defined on the filter coefficients of the model, our approach is inducing sparsity among the weights of the network. Thus, our method is redistributing the information between the model filters, enabling us to remove the filters with the smaller values. We show the performance of our method on classic classification tasks. We are also introducing the efficiency of our technique on more complex models and tasks such as classification, segmentation and detection problems while specialising these models on a subset of categories on several databases.