Conditional DCGAN for Targeted Generation of MNIST Handwritten Digits

Creating accurate images of handwritten numbers is a challenging task that has plagued the field of computer vision for years. Traditional neural networks, such as artificial neural networks (ANN) and convolutional neural networks (CNN), have been unable to accurately generate images that mimic the samples in the training set. However, the Generative Adversarial Network (GAN) has shown promising results in generating images that resemble the training set samples by randomly selecting from the latent space. Despite the effectiveness of GANs, the output of GANs is unpredictable, and it can be challenging to regulate them. To address this, Conditional GANs (cGAN) have been introduced, where additional information can be provided to guide the image generation process. The cGAN enables us to adjust the output of GANs by incorporating additional information from the labels. In this study, the cDCGAN approach to train the MNIST dataset along with the labels was used; cDCGAN stands for conditional deep convolutional GAN, where deep convolutional layers help in extracting more features, and the features extracted can be restored more prominently compared to other GAN architecture like Fully Connected GAN, WGAN, etc. In this research, the training of the cGAN model is done with “10” labels, i.e., 0–9, and the authors intend to do further research in the generation of small case cursive handwriting where there are “26” labels of English letters for which the authors need new dataset of handwritten letters.