A novel multi-loss-based deep adversarial network for handling challenging cases in semi-supervised image semantic segmentation

Image semantic segmentation is popular in computer vision and pattern recognition, since the high-level semantic understanding of images can be effectively realized. Based on whether and to what extent the training data should be labeled, most image semantic segmentation methods can be categorized into fully-supervised learning-based methods, weakly-supervised learning-based methods, and semi-supervised learning-based methods. Among them, semi-supervised image semantic segmentation receives increasing popularity recently, because of its flexibility and convenience in requiring partial training data to be labeled. Although semi-supervised image semantic segmentation is promising, its state-of-the-arts haven't obtained satisfactory performance when handling challenging cases, including poor illumination, small-sized targets, multi-targets with the same semantics, etc. To tackle the above dilemmas, a novel multi-loss-based deep adversarial network is proposed in this paper. Technically, the more robust WGAN-GP model is utilized as the backbone of the novel network, instead of the conventional GAN model. Moreover, multiple losses including the cross entropy loss, the edge detection loss, the adversarial loss, and the semi-supervised loss, are all incorporated during the novel network's training. Experimental analyses based on challenging cases shortlisted from the Pascal VOC 2012 dataset and the Cityscapes dataset suggest that, the novel network is capable to outperform state-of-the-arts in semi-supervised image semantic segmentation.