In the past few years, Generative Adversarial Network (GAN) became aprevalent research topic. By defining two convolutional neural networks(G-Network and D-Network) and introducing an adversarial procedure between themduring the training process, GAN has ability to generate good quality imagesthat look like natural images from a random vector. Besides image generation,GAN may have potential to deal with wide range of real world problems. In thispaper, we follow the basic idea of GAN and propose a novel model for imagesaliency detection, which is called Supervised Adversarial Networks (SAN).Specifically, SAN also trains two models simultaneously: the G-Network takesnatural images as inputs and generates corresponding saliency maps (syntheticsaliency maps), and the D-Network is trained to determine whether one sample isa synthetic saliency map or ground-truth saliency map. However, different fromGAN, the proposed method uses fully supervised learning to learn both G-Networkand D-Network by applying class labels of the training set. Moreover, a novelkind of layer call conv-comparison layer is introduced into the D-Network tofurther improve the saliency performance by forcing the high-level feature ofsynthetic saliency maps and ground-truthes as similar as possible. Experimentalresults on Pascal VOC 2012 database show that the SAN model can generate highquality saliency maps for many complicate natural images.
展开▼
