摘要
传统的步态识别方法难以得到有效的步态特征,而深度学习方法可以通过学习自动获得特征,然而现有的深度学习模型用于步态识别时存在一些问题。深度卷积神经网络训练速度快,但训练精度较低;深度置信网络模型精度较高,但模型收敛速度较慢。针对这两种模型的特点,提出一种两者平衡的算法模型,即深度卷积限制玻尔兹曼机。将卷积神经网络中权值共享、提取图像局部特征等方面的优势融入深度玻尔兹曼机模型中,提高训练精度,减少参数数量。所提算法在CASIA步态数据库上的实验结果验证了该算法在步态识别问题上的有效性和可行性。
In traditional gait recognition methods,it is difficult to get effective gait features.Deep learning methods can automatically learn features.However,the existing deep learning models have some problems when they are used for gait recognition.Convolutional neural networks are fast in training,but the training accuracy is low.Deep belief network has higher accuracy,but it is slower in convergence.According to the characteristics of the two models,a model was proposed,in which deep convolutional restricted Boltzman machine was introduced to balance CNN and DBN.The model takes the advantage of the CNN using its strategies of weight sharing and extracting local features of images.These two advantages were combined with DBN to reduce the number of parameters while improve the training accuracy.Experimental results on the CASIA gait database demonstrate the effectiveness and feasibility of the proposed algorithm on the gait recognition problem.
出处
《计算机工程与设计》
北大核心
2018年第1期244-248,共5页
Computer Engineering and Design
基金
北京自然科学基金重点基金项目B类(KZ201410011014)
教育部人文社会科学研究规划基金项目(16YJAZH072)
