Current Issue Cover


摘 要
目的AdaptSegNet(域自适应的分割网络)在基于弱监督的语义分割中取得了较好的效果,但是该模型在不同特征层的对抗学习中加入了固定惩罚因子,并且该模型直接用合成数据集GTA5作为源域与目标域CityScapes进行对抗训练,故分割精度仍有待提高。方法 针对该问题,提出了一种新的域自适应图像分割方法,用于城市交通场景的分割。算法首先采用SG-GAN方法(Semantic-aware Grad-GAN)对合成数据集进行风格转换,使新生成的数据集SG-GTA5在颜色、纹理和边缘上更接近真实场景数据集CityScapes,并用该数据集代替AdaptSegNet中的源数据集GTA5;然后在网络的不同特征层的对抗学习中使用自适应的惩罚因子,通过该因子调整特征层的损失值,进而动态更新网络参数。此外,为了提高网络的判别能力,算法在对抗网络的判别器中增加了一层卷积层;最后在CityScapes数据集上进行了验证。结果 由于采用SG-GAN方法对合成数据集GTA5进行了风格转换,使得新生成的数据集在边缘、纹理等信息分布上更接近真实场景数据集,有效地减少了源域与目标域之间差异,降低了训练过程中的对抗损失值,避免模型在反向传播过程中出现梯度爆炸,从而提高了分割了精度;另一方面,提出了在不同特征层的对抗学习中使用自适应的惩罚因子,通过该惩罚因子调整各特征层的损失值,进而动态更新网络参数,优化了对抗网络中生成器和判别器的性能,使网络的学习能力得到增强,进一步提高了模型的分割精度;最后在判别器中增加了一层卷积层,使模型能够更好地学习到高层语义信息,提高了本文模型的网络判别能力。结论 在CityScapes数据集上进行了验证,并与现有的性能较好的弱监督分割模型相比,实验结果表明:本文模型能够更加精细的分割出交通场景中的较复杂目标,在sidewalk、wall、car、building等目标都得到较好的分割精度。
A New Semantic Segmentation Method of Urban scene based on Domain Adaptive

zhangguimei,Pan Guofeng,Liu Jian Xin(Nanchang Hangkong University)

Objective AdaptSegNet(Adaptive Segmentation Network) achieves good results in semantic segmentation based on weak-supervisor. However, the fixed penalty factor is introduced in the model during the adversarial learning of different feature layers, and the model is trained directly based on the synthetic data set GTA5 and the real data set CityScapes which are different in distribution, so the segmentation accuracy needs to be improved. Method To address these problems, an adversarial network is proposed for urban scene segmentation,which is capable of adjusting deep neural network parameters in different layers based on adapt learning rate and domain simultaneously. Firstly, the SG-GAN (Semantic-aware Grad-GAN) method is used to train the synthetic data set of GTA5, which makes the newly generated synthetic data set SG-GTA5 closer to the real scene data set in color,texture and edge, and is suitable to substitute the original data set GTA5 in AdaptSegNet. Then, the adaptive learning rate scheme is adopted to adjust loss value and neural network parameters of multi-layers. In addition, a convolution layer is added into the discriminator part in order to improve the discriminant ability of the GAN network. Finally, the algorithm is validated on the CityScapes dataset and compared with several current popular semi-supervised segmentation ones. Result The segmentation precision is improved through the presented data preprocessing scheme on the synthetic dataset of GTA5 by the SG-GAN model, which makes the synthesized dataset closer to the real scene dataset in terms of edge,texture and other information distribution.The provided data preprocessing method also reduces the adversarial loss effectively, and avoids gradient explosion during the back propagation process. Furthermore, the network’s learning capability is enhanced and the model’s segmentation precision is improved further through the presented adaptive penalty factors, which are applied to different adversarial layers to adjust the loss value of each layer.The penalty factor is able to update network parameters dynamically and optimizes the performance of the generator and discriminator network. Finally, the discrimination capability of the provided model is improved more by the additional convolute layer in the discriminator, which enables the model to learn high layer semantic information. Conclusion Comparing with the existing weak supervised segmentation model with good performance on the Cityscapes dataset, the experimental results show that the model can segment more complex targets in the traffic scene more precisely, such as sidewalk, wall, Car, building and other targets.