SPGAN: Siamese projection Generative Adversarial Networks

Abstract

Noise-to-image synthesis continues to be challenging, despite the application of the advanced loss functions in Generative Adversarial Networks (GANs). The main issue lies in the fact that discriminators employ hard margin classification, which is susceptible to misclassification. Moreover, the features of real image distribution learned by the generator are limited during adversarial training, thus the generated images are visually inferior to the real images. To tackle these challenges, a GAN method based on Siamese projection network (abbreviated as SPGAN) is proposed to learn a similarity measurement of features for image synthesis. Then, the similarity measurement is incorporated into the loss functions of the generator and discriminator, forming a similar feature adversarial learning. Through similar feature adversarial learning, SPGAN encourages the discriminator to maximize the dissimilarity between the features of real and generated images during recognition process. Simultaneously, it encourages the generator to synthesize images that contain more features resembling those of real images. Furthermore, we extend the SPGAN method by rewriting five representative loss functions, showcasing its compatibility with different loss functions. Experimental results demonstrate that the performance of SPGAN outperforms the advanced loss functions.