In recent years, research in the field of generative adversarial networks has grown significantly. These networks have various applications such as machine vision and language processing, which have performed amazingly. Among many of these applications, image synthesis has been well studied and researched. The research conducted in the field of image synthesis has shown the high capacity of these networks. The generative model has the ability to generate new data and the training of these models is mainly classified into supervised, unsupervised and semi-supervised. If the generative model can separate the source of variation in the images, it can learn the disentangled representation. In this report, we examine important generative models in the application of image synthesis which have a good disentangled representation and get to know how each of them works. SCGAN is one such generative model that is trained in an unsupervised manner. The difference between the SCGAN and the adversarial generative networks is that in the loss function of SCGAN, a similarity constraint is used as a regularization term and it has gained the ability to learn the disentangled representation based on the similarity between the generated images. We examine the SCGAN in details and understand how it works at a deeper level. This understanding makes us realize that the similarity constraint used, functions like contrastive loss function. So, we get to know the problem of similarity learning and we review some research done in this field. Therefore, we investigate some similarity eva‎luation methods and criteria such as the SSIM. We use this criterion in SCGAN’s similarity constraint because we believe that a model with high understanding and intelligence, measures the similarity between images based on their structure and high-level features, just like humans do. We apply some other changes to the SCGAN to improve its performance and finally implement the modified model. SCGAN obtained the value of the log-likelihood using the Gaussian Parzen window method to measure the performance of the model. This value gives statistical information about the distribution of produced images, but it does not give us information about the variety of images and the power of disentanglement of the learned representation. In fact, the model may suffer from mode collapse, but we observe a good likelihood value. Therefore, we used the appropriate criteria to check the quality of the images, the variety of the images and the power of disentanglement of the learned representation. The results show that the modified model performed better than any other model. Compared to the SCGAN, the log-likelihood of the modified model achieved 234.8 and 332.6 from 232.5 and 324.2 on the MNIST and Fashion-MNIST datasets, respectively. Also, with the FID criterion on these datasets, it has improved from 4.11 and 14.63 to 3.42 and 12.97, respectively. In terms of disentanglement, it improved from 0.77 to 0.89 for MNIST and from 0.89 to 0.91 for Fashion-MNIST and learned a better disenangled representation compared to SCGAN. The modified model has better generalisability than other generative models.

محمدعلي خسروي فرد , محمدرضا حيدرپور

محمدرضا احمدزاده , حامد نريماني