基于空洞卷积鉴别器的语义分割迁移算法

Semantic Segmentation Transfer Algorithm Based on Atrous Convolution Discriminator

近年来,基于卷积神经网络的有监督图像语义分割方法的研究取得了巨大进展。针对该方法所依赖的手动标签繁琐、费时的问题,一种流行的解决方法是通过游戏视频来收集类似于真实场景的图像并自动生成标签,随后利用迁移学习将合成场景训练的模型迁移到真实场景。由于域偏移,简单地将合成场景(源域)上学习的模型应用到真实场景(目标域)一般会出现较高的泛化误差。针对该问题,提出一种新的图像语义分割的无监督迁移算法。该算法首先基于传统的图像风格转换网络对源域图像集进行风格转换预处理,使得图像风格能对齐于目标域,有效降低域间差异;然后,采用生成对抗训练实现源域与目标域特征的对齐。针对现有生成对抗训练中鉴别网络视野受限的问题,提出通过空洞卷积来设计鉴别网络,从而有效提升鉴别网络的分辨能力。在两个典型城市道路数据集GTA5以及SYNTHIA上的实验表明:相比于经典的AdaptSegNet算法,所提算法在GTA5数据集上的平均交并比(mIoU)提高了4.5%,在SYNTHIA数据集上的平均交并比提高了2.6%。

Supervised semantic segmentation with convolutional neural networks has made great progress in recent years.Since the pix-level labeling required by supervised sematic segmentation is tedious and labor intensive,one way that becomes recently prevalent is to collect photo-realistic synthetic data from video games,where pixel-level annotation can be automatically generated.Despite this,the intrinsic domain difference between synthetic and real images usually causes a significant performance drop when applying the learned model to real world scenarios.To solve this problem,we propose a novel domain adaptive semantic segmentation method.It firstly performs image style conversion over the source domain for reducing the domain difference.Then,the generative adversarial network is employed for feature alignment between source and target domains.In particular,we propose to use the atrous convolution for constructing the powerful discriminator network with the enlarged field of view.Extensive experiments show that the proposed algorithm can achieve 4.5%mIoU improvement on the GTA5 dataset and 2.6%on the SYNTHIA dataset,compared with the classic AdaptSegNet algorithm.