基于内容和时空特征的面瘫程度定量评价的鉴别研究

发布时间：2018-07-27 17:00

【摘要】：面部麻痹(面瘫)是一种常见的临床疾病,据统计每10万人中就有20至25名患者。此疾病通常由神经损伤导致,如早期没到得到正确的治疗,可能会留下严重的后遗症。如有一种能有效地客观定量评价患者面部麻痹程度的方法,对治疗是非常有帮助的。针对这种需求,本文应用面部表情数据提取更精确特征信息,为医生更客观地鉴别面瘫患者的面瘫程度提供指导,主要研究成果如下:(1)针对传统研究方法对静态图像因定位、光照、阴影和本身造成不对称性的弊端,基于图像的内容一方面提出了基于LBP和Gabor滤波器相结合的方法(LBP+Gabor)提取不对称特征和运动特征。利用LBP旋转不变性和灰度不变性的优点去除光照影响;再利用Gabor滤波器减少LBP处理后图像中的噪声和冗余信息。另一方面,针对Gabor滤波器中通带的不均匀,提出了基于CMF(Concentric Modulation Filter)方法的特征提取,并将LBP和CMF结合(LBP+CMF)。由于CMF的通带对所有方向都是相同的,能提取更精确的特征。实验表明,LBP+Gabor和LBP+CMF方法相对于传统方法具有较好的有效性和实用性。(2)静态图像中的不相关区域(如嘴巴里的牙齿、阴影和嘴角的痣)影响了特征提取的精确度。为了解决这些问题,本文根据面部数据的数据特征信息,首次提出了应用目标跟踪方法提取时空特征。此方法首先从面部图像第一帧中选取marker,然后在图像序列中追踪marker提取时空特征。本文应用KLT和Mean-Shift两种跟踪方法证明了提取时空特征对鉴别有效性。(3)针对不同患者表情执行的不同步问题,本文对表情的开始时间点和结束时间点进行同步和速度归一化。多类数据实验表明,提取时空特征,并进行同步和归一化后,相比基于内容的方法预测面瘫倒的程度正确率更高,对医生具有更好的实用性。同时,本文提出的方法拓展了目标跟踪方法的应用领域,也为研究面瘫程度定量评估提供了新思路。实验结论有助于医生对患者做出合理的病情判断和治疗。
[Abstract]:Facial paralysis (facial paralysis) is a common clinical disease, according to statistics every 100000 people with 20 to 25 patients. The disease is usually caused by nerve damage, which can lead to severe sequelae if not properly treated at an early stage. If there is an effective method to evaluate the degree of facial paralysis objectively and quantitatively, it is very helpful for the treatment. In response to this demand, this paper uses facial expression data to extract more accurate feature information, and provides guidance for doctors to identify the degree of facial paralysis more objectively. The main research results are as follows: (1) aiming at the traditional research methods to locate the static image, Illumination, shadow and itself cause asymmetry. On the one hand, based on image content, a method based on LBP and Gabor filter is proposed to extract asymmetric features and motion features by (LBP Gabor). The advantages of LBP rotation invariance and gray invariance are used to remove the illumination effect, and the Gabor filter is used to reduce the noise and redundant information in the image processed by LBP. On the other hand, aiming at the nonuniformity of passband in Gabor filter, the feature extraction based on CMF (Concentric Modulation Filter) method is proposed, and LBP and CMF are combined with (LBP CMF). Because the passband of CMF is the same in all directions, it can extract more accurate features. The experimental results show that the LBP Gabor and LBP CMF methods are more effective and practical than the traditional methods. (2) the accuracy of feature extraction is affected by the uncorrelated regions in the still images (such as teeth in the mouth, shadows and nevus in the corners of the mouth). In order to solve these problems, based on the feature information of facial data, a target tracking method is proposed to extract temporal and spatial features for the first time. The method firstly selects Marker from the first frame of facial image, and then tracks the marker to extract temporal and spatial features in the image sequence. In this paper, two tracking methods, KLT and Mean-Shift, are used to prove the effectiveness of extracting spatio-temporal features for discriminating. (3) aiming at the asynchronous performance of different patient's facial expressions, this paper synchronizes and normalizes the start and end points of facial expressions. The experiments of multi-class data show that compared with content-based methods, the accuracy of temporal and spatial feature extraction, synchronization and normalization is higher than that of content-based method, and it is more practical for doctors. At the same time, the proposed method extends the application field of target tracking method and provides a new idea for quantitative evaluation of facial paralysis degree. The conclusion of the experiment is helpful for doctors to make reasonable judgment and treatment of the patient's condition.
【学位授予单位】：中南林业科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R745.12;TN713

【相似文献】