基于生成对抗网络的语义分割方法

A method by Generative Adversarial Network in semantic segmentation

摘要:

在语义分割模型结构不变的前提下，为提升模型对图像分割的精确度，引入生成对抗网络结构用于训练语义分割模型(SS-GAN)。SS-GAN包含3个设计环节：构建全卷积网络(FCN)结构的生成模型，进行初步的图像分割；设计具备像素间高阶关系学习能力的对抗模型，提高生成模型的学习能力；加入对抗损失辅助生成模型学习，进一步促进生成网络自主学习像素间关系。在计算机视觉竞赛数据集(PASCAL VOC)和城市景观数据集(Cityscapes)上的实验结果表明，引入生成对抗网络后取得了更好的效果，2个数据集的交并比(IoU)指标分别提高了1.56%/1.17%和1.93%/1.55%。

Abstract:

In order to improve the accuracy of image segmentation without changing the structure of original semantic segmentation models, an approach is proposed to train Semantic Segmentation models by using Generative Adversarial Network(SS-GAN). There are three steps related to this work: constructing the generative model of Fully Convolutional Network(FCN) structure to segment image preliminarily; constructing the adversarial model which can learn the high-order relationship between pixels and training it to improve the learning ability of generative model; adding the anti-loss to assist generative model training, encouraging generative network to learn the relationship between pixels independently. Experiments on Pattern Analysis, Statistical Modeling and Computational Learning (PASCAL VOC) and Cityscapes datasets show that the proposed method achieves better performance than the existing advanced methods, and improves Intersection over Union(IoU) by 1.56%/1.17% and 1.93%/1.55%, respectively.