基于深度学习的手势识别方法研究

发布时间：2018-05-18 14:44

本文选题：手势识别 + 二值化网络　；参考：《湖南工业大学》2017年硕士论文

【摘要】：手势识别是人机交互一个重要的研究课题,由于对它的研究特别是对基于视觉的手势识别的研究顺应了近年来人机交互从机器友好型向着人类友好型发展的趋势,因此有着极大的科研和应用前景。然而在实际使用中,人手形态的多样性,及其所处环境的背景、光线的变化等因素都给计算机从图像信息中正确识别的手势带来了极大挑战。针对这些问题,本文分别对手势识别、手势检测等问题进行了研究,主要工作如下:(1)针对手势检测问题,结合视频中的多种检测算法提出了一种多策略融合的手势检测方法。为了解决复杂背景下手势检测出现的误检问题,研究了肤色检测、vibe运动检测等算法的原理,根据各种算法在检测中的特点在将肤色、运动和人脸信息进行融合,提升了在复杂背景下手势检测的鲁棒性。特别的针对手势与类肤色背景重合时的检测容易失效问题,对融合策略进行了自适应阈值的改进,改善了算法在该种情况下的检测率。(2)针对手势分类识别问题,在普通的深度学习卷积神经网络手势识别方法的基础上提出了一种基于二值卷积神经网络的手势识别方法。该方法将网络的二值化方法与卷积神经网络手势识别方法相结合,使用二值化后的权值提替代网络中原本的高精度权值,减少了算法计算量及内存占用。通过实验证明,算法在取得了足够的准确性和鲁棒性的基础上,计算效率和在实时系统中的适用性得到了提升。(3)设计和实现了一个手势识别系统,展示了手势识别在人机交互系统中的应用。从系统的需求和功能模块的设计,到结合了前面提出的两种方法的复杂背景下的手势识别功能模块及手势训练模块的实现,再到将成熟的人脸识别检测方案集成的协同认证模块的实现,本文详细地介绍了系统设计实现的各个细节。最后通过实验展示了系统用于识别数字和解锁的功能和特性。
[Abstract]:Gesture recognition is an important research topic in human-computer interaction. Because of its research, especially the research on visual gesture recognition, it conforms to the trend of human-computer interaction from machine-friendly to human-friendly in recent years. Therefore, there is a great prospect of scientific research and application. However, in practical use, the diversity of the human hand shape, the background of the environment and the change of light bring great challenges to the correct recognition of hand gestures from the image information by the computer. Aiming at these problems, this paper studies the problems of gesture recognition and gesture detection respectively. The main work is as follows: (1) aiming at the problem of hand gesture detection, a multi-strategy fusion method for gesture detection is proposed in combination with a variety of video detection algorithms. In order to solve the problem of false detection in hand gesture detection in complex background, the principle of skin color detection and motion detection is studied. According to the characteristics of the algorithms in detection, the color, motion and face information are fused. The robustness of hand gesture detection in complex background is improved. Especially, aiming at the problem that the detection is easy to fail when the gesture and the similar skin color background coincide, the adaptive threshold of the fusion strategy is improved, and the detection rate of the algorithm in this case is improved. On the basis of common deep learning convolution neural network gesture recognition method, a gesture recognition method based on binary convolution neural network is proposed. This method combines the binarization method of the network with the hand gesture recognition method of convolution neural network, and uses the binary weight value to replace the original high precision weight value in the network, thus reducing the computational complexity and memory footprint of the algorithm. It is proved by experiments that the algorithm has achieved enough accuracy and robustness, and the computational efficiency and applicability in real-time system have been improved. (3) A hand gesture recognition system is designed and implemented. The application of gesture recognition in human-computer interaction system is demonstrated. From the design of the system requirements and function modules to the implementation of the hand gesture recognition function module and the gesture training module under the complex background of the two methods mentioned above, Then to the implementation of the collaborative authentication module which integrates the mature face detection scheme, this paper introduces the details of the system design and implementation in detail. Finally, the functions and characteristics of the system for identifying numbers and unlocking are demonstrated through experiments.
【学位授予单位】：湖南工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】