Image-To-Image Translation Using Deep Convolutional GANs

The paper proposes a framework to build a deep neural network (DNN), established by heaping layers of GANs. The GAN architecture comprises both generator and discriminator layers which are locally trained to generate fake images which are of similar resolution as the original input image. The generated inputs are reconstructed exactly as the original input. The algorithm is a candid variation by stacking the ordinary GAN algorithm. The discriminator network is basically a classification problem of machine learning yielding less classification error. Hence, spanning the performance gap with DNNs and in many of the cases surpassing it results in the reconstruction of the inputs depending upon the training parameters such as the upsurge of the epoch and batch size. It will increase the training period, thus increasing the accuracy of synthesized images. A novel mode of training the GANs is discussed in the proposed model. The key point is to grow generator and discriminator progressively, with a low resolution, adding new layers model increasingly and finding details in the model as training progresses. It speeds up training process and also stabilizes results, allowing the results to be of high quality. Additionally, a non-random noise is added and the noise is fixed for the latent space size which yields stable results. A novel metric for evaluating GANs in terms of image quality and variation is discussed.