Hybrid deep learning and swarm intelligence for nanotechnology applications / Dalia Ezzat Aboelyazeed Ali ; Supervised Aboulella Hassanien , Mohamed Hamed

Thesis (M.Sc.) - Cairo University - Faculty of Computers and Artificial Intelligence - Department of Information Technology

Image classification has become a necessary function in nanoscience. Due to the tremendous progress in the techniques used to obtain images in this science, which in turn produced a large number of images in the least possible time such as scanning electron microscope (SEM) technique.The main objective of the thesis is to automatically classify images that are produced from SEM technique using one of the deep learning (DL) architectures.The convolutional neural network (CNN), which is one of the most prominent methods used in computer vision tasks, especially image classification task. Despite the efficiency of the CNN architecture, its success depends on the correct determination of the values of its hyperparameters, such as the rest of the DL architectures. Despite the great importance of the hyperparameters of the CNN architectures, there is no mathematical formula for choosing their optimal values.The values of hyperparameters are usually chosen using the method of manual search, random search method, or grid search which is very time-consuming. In this thesis, a novel approach based on the meta-heuristic algorithms was proposed to determine the best values for the hyperparameters of the CNN architecture used. In order to achieve a high level of accuracy in the classification of images produced by SEM technique.The proposed approach consists of four main phases are (1) data preparation phase, (2) the hyperparameters optimization phase, (3) the learning phase, and (4) the evaluation phase. In the first phase, the used dataset was addressed from the imbalance problem using two simple but effective methods, which are: random oversampling (ROS) and random undersampling (RUS)

Issued also as CD