Deep learning is a subset of Machine Learning that is based on artistic neural networks that attempt to simulate the behavior of the human brain, or biological neural networks. Compared to conventional Machine Learning methods, these methods are very good at dealing with data with a large number of dimensions or with non-structured data such as images, video, text and sound. This Thesis is based on a special case of data reconstruction in the form of an image and such problem is part of computer vision as an area of deep learning. The deep learning network architectures used are the Convolutional Neural Network (CNN) and special Generative Adversarial Networks (GAN) that are based on Nash’s equilibrium as a learning concept. Specifically, the Auto – Encoder Generative Adversarial Network (AEGAN) and the Context – Conditional Generative Adversarial Network (CCGAN). The differences in the mentioned architectures are presented and the results obtained in the reconstructions, ie the completion of the lost information in the images, are compared. The dataset used to train these networks is a publicly accessible set of images of human faces (CelebA). The Thesis implementation was done with the Python programming language and the PyTorch deep learning programming library. The results obtained in the Thesis show the following; the CCGAN method has proven to be the best in the case where a large part in image is unknown and could be used for such applications. In the case of random unknown pixels, the CNN and AEGAN methods showed very similar results, but taking into account the metrics of reconstruction accuracy and the complexity of architecture and implementation, following the principle of „Occam’s razor“, CNN is taken as the best presented method in these cases masking. This Thesis shows the success of the methods in the application of data completion and reconstruction that can serve as a basis for further work on real world and industrial applications, such as geosciences, oceanography and satellite applications where the number of measurements in space is limited and would result in reduced consumption due to reduced number of sensors and measurements.