Generative Adversarial Networks (GAN’s) are becoming effective process of using Artificial Intelligence technology for the system containing of two neural networks to compete each other. This framework in terms containing both generator and discriminator models and places a game type competition in between them. Now the Generator model(G) and the Discriminator model(D) will be simultaneously trains and generates the data that is distributed and estimates the probability for it. And it can be very interesting to understand the GAN and deep convolutional neural networks models are trained for the digits or the image generations. By the help of MNIST database the system can be processed to rectify the real images which are present in the loaded database and compares it to find a match.
We have conducted the project with the help of MNIST database and we used the handwritten digit based database to test the architecture of generated output and considered the usage of numbers from 0 to 9 for this purpose. We train the GAN with respect to the required amount of epochs and batch normalization to gain the realistic images, given below are the figures which represents the step wise process and whenever the training.
So from these results we can find that DCGAN uses the convolutional neural networks into the GAN architecture and uses the better parameters and convolutional layers to improve the performance of GAN and can generate the realistic images in the less training time too. Apart from the GAN the DCGAN is the present evolving for all types of image segmentation based problems and to classify these type of handwritten digits in a clearer way and we know the training process is time taking but regarding to the outputs it is better and also the generated image process is similar to the GAN.