Deep Neural Networks for Image Restoration in MNIST Dataset
In our work we try to present concisely and comprehensively the basic features of the structure of a multiple layer neural network and some fundamental technics and methodologies that determine its function. Concerning the fact that some chapters are quite tricky, our approach is more intuitive in order to be explained and understood better. At the same time, taking into account that neural networks and Machine Learning are based on strictly mathematical models, we focus on mathematic explanation wherever is essential.
Our goal is the initiation of a new reader to Deep Learning, regardless of his elder knowledge of this subject. After reading this thesis he will be able to understand how Deep Learning works and its general applications. Although in our work we have included beyond the basic techniques and several modern ones that are used even in today's applications, we would suggest to someone who is interested in studying further, to look for the new trends and methods that are more specialized and up to date.
Considering that Neural Networks are one of the most basic and effective image processing methods, we experimented with several types of Neural Networks in order to reconstruct corrupted images.
We developed two kinds of Neural Networks. More specifically, we implemented a Convolutional Neural Network (CNN) and an Encoder - Decoder to solve both Image Inpainting and Image Denoising. We used MNIST Dataset which consists of tiny images that illustrate handwritten digits in range 0 to 9. This Dataset is the most suitable as it facilitates our experiments with the networks' parameters and it does not require an expensive machine to be used.
One of our goals is to solve these problems and to produce high quality images. We also aim to draw conclusions by experimenting with both networks' and problems' parameters.
Our application has been developed using Tensorflow library.