复杂背景下的静态手语识别技术研究

发布时间：2018-04-22 02:19

本文选题：肤色聚类 + 图割　；参考：《沈阳工业大学》2017年硕士论文

【摘要】：进入21世纪以来,人机交互系统成为了科学界的宠儿,受到了一大批科研学者的青睐,其中手语识别技术的研究炙手可热。随着研究的深入以及社会需求的提高对手语识别的要求更加苛刻,基于视觉的手语识别成为当下研究的重点。传统的手语识别方法对环境背景,图像采集设备以及人的手势动作要求都较高,因此针对手语图像背景的复杂性,手语手形的相似性以及手语识别算法的局限性,本文提出一种新的静态手语识别系统。本文针对手语图像背景的复杂性,提出一种基于改进的肤色聚类的初分割以及利用图割方法的再分割算法,通过该方法可以在复杂背景中获取完整的手区域;其次针对手形的相似性,提出利用链码跟踪方法检测手语图像的轮廓,随后提取轮廓的凸包缺陷点、Hu矩、几何特征以及区域描绘子等多个特征值用来区分各种手语,提取的这些多特征具有综合性和区分性,能够很好地体现各种手语的形状特征;最后针对手语种类的多样性以及手语识别算法的局限性,提出多次分类识别算法即采用SVM进行依据缺陷点个数的初分类以及利用多特征融合的再分类识别算法,之后针对存在一些难以区分且非常相似的手语利用两个区域描绘子进行细分类识别,通过对手语图像进行多次的分类识别可以更好地区分每种手语且运算量小。实验得出在自采集的26种静态手语图像分类识别中平均识别率达93.18%,表明了本文提出的手语识别系统的有效性和社会实用性。
[Abstract]:Since the beginning of the 21st century, the human-computer interaction system has become the favorite of the scientific community, and has been favored by a large number of researchers, among which the research of sign language recognition technology is very hot. With the deepening of research and the improvement of social needs, sign language recognition based on vision has become the focus of current research. Traditional sign language recognition methods require high environmental background, image acquisition equipment and gesture action. Therefore, the complexity of sign language image background, the similarity of sign language hand shape and the limitation of sign language recognition algorithm are discussed. This paper presents a new static sign language recognition system. Aiming at the complexity of the background of sign language image, this paper proposes a new algorithm of initial segmentation based on improved skin color clustering and a resegmentation algorithm using graph cutting method, through which the complete hand region can be obtained in the complex background. Secondly, aiming at the similarity of hand shape, a chain-code tracking method is proposed to detect the contour of sign language images, and then to extract the Hu moments, geometric features and region descriptors of the contour to distinguish various sign languages. The extracted features are comprehensive and distinguishable, which can well reflect the shape features of various sign languages. Finally, aiming at the diversity of sign language types and the limitations of sign language recognition algorithms, In this paper, a multiple classification recognition algorithm is proposed, that is, the initial classification based on the number of defect points and the subclassification recognition algorithm based on multi-feature fusion are proposed by using SVM. Then we use two region descriptors to subdivide the sign language which is difficult to distinguish and very similar. By classifying and recognizing the sign language image for several times we can distinguish each sign language better and the computation is less. The experimental results show that the average recognition rate of 26 static sign language images collected by ourselves is 93.18, which shows the effectiveness and social practicability of the sign language recognition system proposed in this paper.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】