手势图像识别算法研究

发布时间：2018-06-05 11:41

本文选题：手势识别 + Otsu分割算法　；参考：《沈阳理工大学》2017年硕士论文

【摘要】：人与计算机的频繁互动已经成为生活中的日常操作,其中,关于手势的研究已然成为目前人机交互研究领域的主要研究方向之一。手势识别技术的研究将会改变传统的人机交互方式,手势的的使用必然会使得人机交互技术从以机器为中心逐步转移到以人为中心,让人机交互的方式变得便捷,人机交互方式变得丰富,也使使用计算机的门槛降低。本文对手势识别完整系统进行表述,主要包含了四个主要部分,静态手势图像的预处理,关于手势的图像分割,手势的特征提取以及最后的手势的识别方法。系统通过摄像头捕获手势图像,对该图像进行预处理,其中包括彩色空间转换、平滑处理、形态学运算、灰度化、二值化、轮廓提取,其中详细介绍通过常用的颜色空间,分析影响到手势特征提取及分割的色彩分量,并通过彩色空间转换减弱甚至消除该影响。本文详细介绍基于canny边缘的检测方法,并根据在手势边缘提取方法上的不足提出改进。手势分割部分是手势识别系统的关键步骤之一,在简单、单一背景的室内环境下分割手势的算法要求不高,但是在复杂背景下的室外环境下,有太多的干扰,这使得传统的分割方法无法将手势从背景中干净的分割出来,本文介绍的传统的Otsu算法虽然在单一背景下效果不错,但是复杂背景下显得捉襟见肘,通过改进灰度图像的划分方法使的Otsu算法能够分割出手势。干净的手势图像中的信息量太多,如果作为分类识别系统的输入,这会增加识别系统的计算量以及计算复杂度,所以手势图像的特征提取是需要的,本文使用的是图像区域几何的特征的不变距,不变距由7个不变距的值组成,我们为了使识别系统分类的输入具有旋转、平移、尺度变化不变性,就需要通过仿真并且比较从中挑选出符合条件的分量并组合成输入向量。在识别方法的挑选中,本文挑选的基于自适应神经-模糊推理系统(ANFIS)的手势识别方法具有自主学习的能力,而且鲁棒性好。虽然该方法的识别能力好但是计算复杂度高,我们通过对不变距的筛选结合自适应神经-模糊推理系统的手势识别法,提高整个系统的手势识别率,并且与BP神经网络和模糊神经网络进行,平均识别率95.3%说明自适应神经-模糊推理系统在识别率方面的效果,符合高识别率的实际准则。
[Abstract]:The frequent interaction between human and computer has become a daily operation in daily life, among which, the research on gesture has become one of the main research directions in the field of human-computer interaction. The research of gesture recognition technology will change the traditional human-computer interaction mode. The use of gesture will inevitably make the human-computer interaction technology shift from machine center to human-centered, so that the human-computer interaction becomes convenient. The man-machine interaction way becomes rich, also causes the computer to use the threshold to lower. This paper describes the complete system of gesture recognition, which includes four main parts: the preprocessing of static gesture image, the segmentation of gesture image, the feature extraction of gesture and the final gesture recognition method. The system captures the gesture image through the camera and preprocesses the image, including color space conversion, smoothing processing, morphological operation, grayscale, binarization, contour extraction, in which the commonly used color space is introduced in detail. The color components which affect gesture feature extraction and segmentation are analyzed, and the influence is weakened or even eliminated by color space conversion. This paper introduces the method of edge detection based on canny in detail, and proposes some improvements according to the shortcomings of the method of gesture edge detection. Gesture segmentation is one of the key steps of gesture recognition system. The algorithm of hand gesture segmentation in simple, single background indoor environment is not high, but in the outdoor environment of complex background, there is too much interference. This makes it impossible for traditional segmentation methods to segment gestures from the background cleanly. Although the traditional Otsu algorithm introduced in this paper has a good effect in a single background, it appears to be overstretched in a complex background. By improving the grayscale image partition method, the Otsu algorithm can segment the gesture. There is too much information in a clean gesture image. If it is used as the input of the classification recognition system, it will increase the computation and complexity of the recognition system, so the feature extraction of the gesture image is needed. In this paper, we use the invariant distance of the geometric features of the image region. The invariance is composed of seven invariant values. In order to make the input of the classification of the recognition system have the invariance of rotation, translation and scale change. It is necessary to select the suitable components and combine them into input vectors by simulation and comparison. In the selection of recognition methods, the gesture recognition method selected in this paper based on adaptive neural fuzzy inference system (ANFIS) has the ability of autonomous learning and good robustness. Although this method has good recognition ability and high computational complexity, we improve the gesture recognition rate of the whole system by selecting invariant distance and combining with the gesture recognition method of adaptive neural fuzzy inference system. Compared with BP neural network and fuzzy neural network, the average recognition rate is 95.3%, which shows that the adaptive neural fuzzy inference system is effective in recognition rate and accords with the practical criterion of high recognition rate.
【学位授予单位】：沈阳理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】