毕业生知识库

近年来，随着立体显示技术的快速发展，3D电视等立体显示应用已经进入了大众的生活。相比普通的平面显示技术，3D显示可以提供更加真实的体验。然而随着立体显示技术的应用范围越来越广，它本身的成像原理与人体视觉系统的冲突所带来的视疲劳等问题也日益突出，并且逐渐制约了立体显示技术的发展。由于目前技术水平的限制，长期观看立体显示内容导致的视疲劳等健康问题还无法消除，因此需要通过研究立体显示视疲劳评估方法，对用户的视疲劳状态进行合理、准确的评估，从而降低该技术带来的负面影响，提升用户体验，使得立体显示技术得到更加广泛的应用。
本文使用脑电信号作为客观评估指标，将深度学习作为建模方法对视疲劳评估问题进行了研究。本文针对传统的视疲劳评估存在的问题，结合脑电信号的特点提出了一种端到端的视疲劳评估模型，对深度学习方法在视疲劳评估任务中的应用进行了探索。除此之外，本文针对脑电信号分类任务中标记数据不足这一问题进行了研究，并提出了一种基于多任务学习的解决方案。本文的主要工作如下：
1.针对脑电信号特点，提出了一种深度学习模型DeepFatigueNet用于脑电信号分类。该方法受传统脑电信号特征提取算法启发，通过特殊设计的输入卷积模块分别提取脑电信号的时间特征与空间特征并进行分类。整个过程以端到端学习的形式进行，避免了复杂、耗时的传统手工特征提取过程。
2.针对脑电信号分类任务缺乏大规模标记数据这一问题，首次提出了一种基于多任务学习框架的深度学习方法，该方法包含三个部分，分别为表示学习、分类模块以及重建模块。基于多任务学习框架，三个模块以端到端的形式进行优化。本方法可以在标记数据有限的情况下，进一步提升深度学习模型在脑电信号分类任务上的效果。
3.将深度学习应用到视疲劳评估任务中。基于视疲劳评估数据集，将本文提出的DeepFatigueNet模型与传统方法和其他深度学习结构进行对比。实验结果显示DeepFatigueNet的表现超越了其他方法，证明了深度学习在视疲劳评估中的可能性以及DeepFatigueNet的有效性。
4.将多任务学习模型应用到视疲劳评估与运动想象分类任务中。基于视疲劳评估数据集与运动想象公开数据集BCI Competition IV dataset 2a，本文将多任务学习模型与传统方法和state-of-the-art深度学习模型进行了对比。实验结果表明标记数据不足确实限制了深度学习方法表现的提升，而本文提出的多任务学习方法可以在标记数据有限的情况下进一步提升深度学习模型的表现。以上结果证明了多任务学习方法的有效性和通用性。

In recent years, with the rapid development of stereoscopic display technology, stereoscopic display applications such as 3DTV have entered people’s life. Compared with the ordinary 2D display technology, 3D display can provide a more realistic visual experience. However, with the application range of stereoscopic display technology getting wider, the problem of visual fatigue caused by the conflict between its imaging principle and the human visual system is also becoming increasing prominent and gradually restricts the development of 3D display technology. Due to current technical limitations, health problems such visual fatigue caused by viewing 3D content cannot be eliminated. Therefore, it is important to make a reasonable and accurate assessment of the level of the visual fatigue of users to reduce the negative effects brought by the 3D technology and enhance the user experience.
In this paper, EEG signals are adopted as objective assessment indicators and deep learning is used as modeling methods to study the problem of visual fatigue assessment. Aiming at the problem of traditional visual fatigue assessment and considering the characteristics of EEG, an end-to-end visual fatigue assessment method is proposed. The application of deep learning methods in visual fatigue assessment is explored. In addition, problem of insufficient labeled data in the classification task of EEG signals is discussed in this paper and a solution based on multi-task learning is proposed. The key points of this paper are as follows:
1.    Based on the characteristics of EEG signals, a deep learning model DeepFatigueNet is proposed for EEG classification task. The method is inspired by the traditional feature extraction algorithms for EEG signals. The temporal and spatial features of EEG signals are extracted by a specially designed convolutional block and classified. The whole process is carried out in an end-to-end learning manner, avoiding the complicated and time-consuming traditional feature extraction process.
2.    Aiming at the problem of lacking of large-scale labeled data for EEG classification task, a deep learning method based on multi-task learning framework is proposed. This method consists of three parts, namely representation learning module, classification module and reconstruction module. Based on the multi-task learning framework, the three modules are optimized in an end-to-end manner. This method can further improve the performance of deep learning models when there is limited labeled data.
3.    Apply deep learning methods to visual fatigue assessment. Based on the visual fatigue assessment dataset, DeepFatigueNet is compared with traditional machine learning method and other deep learning structures. The experimental results show that DeepFatigueNet outperforms other methods, which demonstrates the potential of deep learning in visual fatigue assessment and the effectiveness of our model.
4.    Apply multi-task learning model to the visual fatigue assessment and motor imagery (MI) classification. Based on the visual fatigue assessment dataset and public dataset BCI Competition IV dataset 2a for MI, the multi-task learning model is compared with state-of-the-art deep learning methods. The experimental results show that the problem of insufficient data does limit the improvement of performance of deep learning methods. Moreover, our proposed multi-task learning method can further boost the performance of deep learning models with limited data. The above results demonstrate that our proposed multi-task learning method is effective and generic.