| Popis: |
GANs or generative adversarial networks are deep generative models that have the capability to imitate a given data distribution and generate realistic data, that can be in the form of images, music, and even outcomes of scientific experiments. An important application of GANs is data creation and augmentation in cases where there is not enough data or the variety in the data is less for training a model to perform well in real-world testing. A lot of applications ranging from license plate recognition, bank check processing, auto filling form details, etc., make use of digit recognition. The digit recognition and classification accuracy of machine learning models, generally if not always, can be improved when more data (images of digits in this case) is provided. It can help the model to learn distinct features from the wide variety of data for better predictions and generalization. In this study, we use GANs and the SVHN dataset of 73 thousand colored, house number images to generate new and unseen, similar, realistic looking images of digits and then evaluate our GAN model using the Inception Score as the metric for evaluating quality and diversity of newly generated images. Later with an intent to further improve the quality of the newly generated images, we study the effects of modifying the discriminator (CNN) architecture by adding residual blocks to make it deeper, along with some other minor experimentation with hyperparameters, with an expectation of improving the classification accuracy of the discriminator, thus helping the generator improvise in generating more realistic images and finally discuss the results. |
