Difficult images with complicated environments and occlusion have significant impacts on the performance of algorithms. They obey the long-tail distribution in the widely used datasets, which results in rare samples being overwhelmed during training. This paper presents a new approach to generate plausible occluded images with annotation as a kind of data augmentation with scenes semantics. To achieve this task, we proposed the Occlusion-based Generative Adversarial Network (OccGAN) structure, which consists of a Rationality Module and an Authenticity Module. The Rationality Module generated preliminary occluded samples under the guidance of prior semantic knowledge. And the Authenticity Module is a generative adversarial structure to ensure the reality of the produced images. Qualitative results of the visualization process are given to verify the ablation study. Experiments on the semantic segmentation task indicate that several state-of-the-art algorithms combined with our OccGAN such as DRN, Deeplabv3+, PSPNet and ResNet-38, have boosts on IoU class scores and IoU category scores successfully.

具有复杂环境和遮挡的困难图像会对算法的性能产生重大影响。他们服从于广泛使用的数据集中的长尾分布，这导致训练期间稀有样本不堪重负。本文提出了一种新的方法来生成带有注释的合理遮挡图像，作为一种具有场景语义的数据增强方法。为了实现此任务，我们提出了一种基于遮挡的生成对抗网络（OccGAN）结构，该结构由合理性模块和真实性模块组成。理性模块在先验语义知识的指导下生成了初步的被遮挡样本。真实性模块是一种生成式对抗结构，可确保所生成图像的真实性。给出了可视化过程的定性结果，以验证消融研究。语义分割任务的实验表明，结合我们的OccGAN的几种最新算法，例如DRN，Deeplabv3 +，PSPNet和ResNet-38，成功地提高了IoU类评分和IoU类别评分。