【作者】 孙欣；

【导师】 张健康； 张勇；

【作者基本信息】 西安电子科技大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 人类很大程度是依靠人眼来获取外部信息,通过对人眼视线点的分析,可以获得人类的感兴趣点,从而得到人类的心里活动信息。由于人眼的重要性,如今已经生产出很多以人眼作为媒介的眼动设备,它在临床医学研究、心理学研究及教育研究等领域,有着广泛的发展前景。虹膜的定位与眼动数据分类是实现眼动设备的关键技术。因此,本文对这两个技术进行了分析与研究,并取得较好的实验结果。以下几点为本文的主要内容与创新点:(1)现有的基于传统算法的虹膜检测鲁棒性较差,而基于深度学习的算法检测效果较好,但其计算量大、检测速度慢。本文针对这一缺陷,提出了一种基于深度学习的轻量化网络结构DDSNet,此网络采用密集块的结构,能够在卷积核个数很少的情况下有着较好检测效果。通过使用深度可分离卷积,进一步的减少了模型的计算量。为了提高模型性能,本文加入了数据增强策略,对训练集的图片进行变换,增大其规模,提高了模型的鲁棒性。在虹膜数据集上对本文的结构进行了评估,相比于其它的轻量化网络,模型的速度显著提高,存储量明显降低。模型在GPU下1秒可以检测639张图片,存储量进仅为1.3M,F1得分达到了99.95%,可以看出本文的模型在速度、存储量与精度上均有着很好的表现。(2)本文使用了基于深度学习的眼动分类算法,此方法首先提取了5个不同时间尺度的特征,将其输入到一维卷积中提取内部特性,并使用BLSTM获取数据集中前后时序间的关系。在Gaze Com眼动数据集上进行测试,发现平滑尾随点的F1得分较低,故通过以下方式对算法进行改进。对少数类样本平滑尾随点与眼跳点采用重采样和人工合成对数据进行增强,提高了少数类样本的数据量。使用了focal loss损失函数,增加了难样本的学习权重,加强对难样本的学习。在网络结构中加入注意力模块,对特征重标定,加强了重要特征。通过这些方式的改进,使模型有了更好的分类能力,使注视点、平滑尾随点与眼跳点的F1得分分别提高了1.2%、3.5%和2.4%。更多还原

【Abstract】 To a large extent,humans rely on the human eye to obtain external information.Through the analysis of the point of view of the human eye,human interest points can be obtained,thereby obtaining information of human mental activities.Due to the importance of the human eye,many eye movement devices that use the human eye as a medium have been produced.It has broad development prospects in the fields of clinical medical research,psychological research,and educational research.Iris positioning and eye movement data classification are the key technologies to realize eye movement equipment.Therefore,this paper analyzes and studies these two technologies and obtains good experimental results.The following points are the main contents and innovations of this article:(1)The existing iris detection based on traditional algorithms is less robust,while the deep learning-based algorithm has better detection results,but its calculation is large and the detection speed is slow.Aiming at this defect,this paper proposes a lightweight network structure DDSNet based on deep learning.This network uses a dense block structure,which can have a good detection effect when the number of convolution kernels is small.By using depth separable convolutions,the calculation of the model is further reduced.In order to improve the performance of the model,a data enhancement strategy is added in this paper to transform the pictures of the training set to increase its size and improve the robustness of the model.The structure of this paper is evaluated on the iris data set.Compared with other lightweight networks,the speed of the model is significantly improved,and the storage capacity is significantly reduced.The model can detect 639 pictures in 1 second under the GPU,the storage capacity is only 1.3M,and the F1 score reaches 99.95%.It can be seen that the model in this article has a good performance in speed,storage and accuracy.(2)In this paper,an eye movement classification algorithm based on deep learning is used.This method first extracts five features of different time scales,inputs them into one-dimensional convolution to extract internal characteristics,and uses BLSTM to obtain the relationship between the time series before and after the data set.After testing on the Gaze Com eye movement data set,it was found that the F1 score of the smooth pursuit point was low,so the algorithm was improved in the following way.The smooth pursuit points and saccade points of minority samples are enhanced by resampling and artificial synthesis to increase the data volume of minority samples.The loss function of focal loss is used to increase the learning weight of difficult samples and strengthen the learning of difficult samples.Attention module is added to the network structure to recalibrate the features and strengthen the important features.Through these improvements,the model has a better classification ability,and the F1 scores of the fixation point,smooth pursuit point and saccade point are increased by 1.2%,3.5% and 2.4%,respectively.更多还原