Handwriting Digit Generation Based on GAN Model

Over the past decades, generating pictures of handwritten digits has always been a challenging task. However, most neural networks do not have the function to imitate the real samples of the training set. By contrast, the Generative Adversarial Network (GAN) can randomly sample from the latent space and output results that mimic the real samples in the training set. In this work, we take advantage of the excellent characteristics of GAN and then realize the generation of handwritten digital pictures through the learning of the MNIST data set. The experimental results demonstrate that the GAN model achieves superior accuracy and robustness of handwritten digits generation. Furthermore, we note that the network parameters could be adjusted constantly through the formation of new adversarial data sets, which could enhance the accuracy of picture identification dramatically.