CASIA OpenIR  > 毕业生  > 硕士学位论文
基于领域自适应的语义分割方法研究
曾鸣
2023-05-20
Pages96
Subtype硕士
Abstract

深度学习技术的发展提高了计算机对图像的感知理解能力,尤其在自动驾驶、医疗影像等领域发挥了重要作用。语义分割作为计算机视觉的重要任务,已成为精细化感知理解应用中的关键部分。然而,在实际应用中,由于训练数据与应用场景的不同分布以及真实场景数据的标注难度,模型在开放世界中的性能往往下降明显,阻碍了分割模型在真实场景中的部署。为了解决上述问题,研究者们提出领域自适应学习范式,利用已有标注的数据集和实际应用场景之间的相关性,使模型在新场景中更好地发挥作用。然而,在领域自适应问题上,目标域上监督信息的缺失和场景的持续变化是两大难题。本文探索了语义分割任务下的无监督领域自适应与持续性领域自适应两大问题,解决了模型在跨领域部署时的性能损失问题,并使模型能够持续适应新场景。其主要创新点如下:

1.提出了一组基于高斯滤波变换的无监督领域自适应语义分割方法,它能够在不依赖对抗学习的情况下,从外观分布和类别分布两个方面对齐领域并缓解差异,避免了额外的训练,有效提高了模型的性能和鲁棒性。目前领域自适应的主流方法之一就是利用对抗学习使模型掌握更多领域不变知识,但容易造成目标领域特定信息的丢失,陷入次优解的困境。因此,本文在基于傅里叶变换的领域自适应方法基础上,引入高斯滤波,并统计(伪)标签的类别分布,在无需对判别器或其他组件进行额外训练的情况下,有效提升了模型在目标域上的性能。

2.提出了一种新的基于参数转移的持续性领域自适应语义分割方法,能有效解决灾难性遗忘问题,为持续性领域自适应语义分割的研究与应用提供了一种新的视角与思路。本文基于假设——领域自适应过程中模型参数可以分为可转移参数(学习领域不变知识)和不可转移参数(学习领域特定知识)——开展研究和探讨,建立可以区分这两类参数的指标,并限制可转移参数在新领域上的更新,成功避免了灾难性遗忘的发生,并且在新领域上表现出良好的性能。

综上所述,本文通过对语义分割任务下的无监督领域自适应与持续性领域自适应两大问题进行深入分析,从外观分布和类别分布以及模型参数性质出发,提出了高斯滤波-傅里叶领域自适应、面向已知类别的重加权、基于参数转移的持续性领域自适应等方法,具有一定的理论价值和应用价值。

Other Abstract

The development of deep learning has improved the computer's ability to understand images, especially playing an important role in fields such as autonomous driving and medical imaging. Semantic segmentation, an important task in computer vision, is significant in applications that perform precise perceptual analysis and understanding. However, due to the difference between training data and real-world scenarios, and the difficulty of labeling data, the performance of the well-trained model often decreases significantly in the open scenarios, hindering the deployment of segmentation models. To address these challenges, researchers have proposed domain adaptation, which utilizes a labeled dataset and real-world scenarios to enhance the model's performance in real-world scenarios. However, the lack of supervised information in the target domain and the continuous change of real-world scenarios are two major problems in domain adaptation. We explore unsupervised domain adaptation semantic segmentation and continual domain adaptation semantic segmentation, to solve the cross-domain performance degradation, and enable the model to adapt to new scenarios continuously. The main contributions include:

1. proposes a set of domain adaptive methods based on the Gaussian Filter and Fourier Transform, which align domains from both the appearance and class distributions without relying on adversarial learning. These methods effectively improve the performance and robustness of the model. One of the mainstream methods in unsupervised domain adaptation is to use adversarial learning to make model learning more domain-invariant knowledge, but it is easy to ignore domain-specific information in the target domain and result in sub-optimal performance. Therefore, based on Fourier Domain Adaptation, we use a Gaussian Filter and count the class distribution of (pseudo) labels, which effectively improves the model's performance in the target domain without additional training for discriminators or other components.

2. proposes a new continual domain adaptive semantic segmentation method based on transferable parameter learning to solve the catastrophic forgetting problem in continual domain adaptive semantic segmentation and provides a new idea for continual domain adaptive semantic segmentation. Based on the hypothesis that the model parameters in domain adaptation can be divided into transferable parameters that are more inclined to learn domain-invariant knowledge and non-transferable parameters that are more inclined to learn domain-specific knowledge, we set an indicator that can distinguish these two types of parameters and limit the transferable parameters' update in the new domain to avoid catastrophic forgetting and achieve good performance in the new domain. 

In summary, we focus on unsupervised domain adaptation and continual domain adaptation for semantic segmentation and propose several methods from the perspective of the appearance and class distribution alignment, as well as the update of model parameters, which have certain theoretical and practical value.

Keyword深度学习 领域自适应 语义分割 持续学习 域增量
Language中文
Sub direction classification机器学习
planning direction of the national heavy laboratory虚实融合与迁移学习
Paper associated data
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/51888
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
曾鸣. 基于领域自适应的语义分割方法研究[D],2023.
Files in This Item:
File Name/Size DocType Version Access License
202028014629013曾鸣.pd(16161KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[曾鸣]'s Articles
Baidu academic
Similar articles in Baidu academic
[曾鸣]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[曾鸣]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.