Stacked Siamese Generative Adversarial Nets: A Novel Way to Enlarge Image Dataset

Deep neural networks often need to be trained with a large number of samples in dataset. When the training dataset are not enough, performance model will degrade. The Generative Adversarial Network (GAN) is considered effective at generating samples, and thus, expanding datasets. Consequently, this paper, we proposed novel method, called Stacked Siamese (SSGAN), for large-scale images high quality. SSGAN made Color Mean Segmentation Encoder (CMS-Encoder) several Networks (SGAN). CMS-Encoder extracts features from using clustering-based method. Therefore, does its output has interpretability human visuals. (SGAN) controls category generated while guaranteeing diversity by introducing supervisor WGAN. progressively learns feature pyramid. We compare Fréchet Inception Distance (FID) previous works on four result shows that our method outperforms works. In addition, CelebA dataset, which consists cropped size 128 × 128. good visual effect further proves outstanding images.