الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 81 |  |  |
| | | from | 81 | | |
Abstract
Deep Learning is a new field in Machine Learning, it helps artificial neural networks to learn multi-level presentation models and multiple non-linear transformations of data for helping layers to simulate the visual cortex in humans, and It makes learning algorithms easier to use about programming models. Deep convolutional neural networks use shared weights to reduce connection between nodes and provide it a large solution space. The target function of deep learning is complex and the dataset is usually large. It has several application such as image classification. In this study, we aims to classify low quality tiny images by using multi-stage model for increasing the total performance of it. Our classification model consists of three main stages, the first stage (preprocessing step) converts RGB images to gray-scale mode and applies ZCA whitening transformation. The second stage focuses on extracting the useful features by using deep neural network model and trains data by deep belief network algorithm that uses Accord machine learning library designed by Microsoft Team for implementing it. The third stage uses the trained data for extracting extra features by using Stochastic Gradient Decent trainer algorithm and it implemented by Tensorflow machine learning library that designed by Google Team. When applying the proposed model on CIFAR-10 dataset, the results shows that it improves the quality of ting images accuracy by 90.45%. The second proposed work for improving STL-10 dataset by using a combination of convolutional neural networks and deep belief network to construct an efficient model that able us to classify low quality images. This model has the capability in extracting effective features from low quality images. Data augmentation is used through this model for increasing the accuracy of the system. Scikit-Learn python library is used in implementation the system on STL-10 dataset. The results showed that the proposed model increase the accuracy of the system by 73.00%.