S$^{{\text{3}}}$Seg: A Three-Stage Unsupervised Foreground and Background Segmentation Network

Abstract

Generative adversarial networks (GANs) gather increasing attention in the field of unsupervised image segmentation. However, most GAN-based unsupervised segmentation methods cannot directly segment a specific individual image, and the segmentation performance can be further boosted. To address these issues, we propose a three-stage unsupervised foreground and background segmentation network (S ³ Seg). In the first stage, we design low-to-high dimensional attention (LHA) to enhance the image understanding ability of the StyleGAN2 generator, which can capture more effective semantic information for the following segmentation task. In the second stage, we introduce an inversion network to form an encoder-generator structure, which can acquire semantic features of a given image. In the third stage, we devise a radial loss to explore the edge of the foreground from the center to the outside, which is beneficial for producing a high-quality mask. S ³ Seg not only provides a solution to direct segmentation of a given image but also outperforms previous unsupervised methods on three public datasets.