摘要
传统的手背静脉身份识别研究对旋转,平移,尺度变化敏感,极大地影响了识别率;因此该文在非限定采集者手背姿势的情况下,将人类视觉注意力机制与神经网络结合,针对旋转,平移,尺度变化等问题提出了优化视觉聚焦点的循环神经网络模型;该模型自适应寻找手背静脉聚焦点,以聚焦点为中心,截取局部ROI区域,送入循环神经网络训练各局部区域的序列关联性;该文的优化方法如下:在选取聚焦点时,加入正态分布噪声;对聚集点的个数进行约束;截取多尺度局部ROI;训练时采用强化学习中的策略梯度下降法和最优化的无偏估计交叉熵损失函数;将该循环神经网络网络模型在多形态的手背静脉数据中进行实验验证,识别率达到99.3%,与传统的局部特征提取方法相比,极大的提高了手背静脉的识别率。
Traditional hand vein identification is sensitive to rotation, translation, scale, which greatly affects the recognition rate. In this paper, in the case of unrestricted hand posture, a recurrent neural network (RNN) of visual attention model is proposed to deal with the problem of variations of rotation and translation. The model adaptively selects the focus of dorsal hand veins, takes the focus as the center crop the local ROI regions, and sends it into the RNN to train the sequence correlation. The optimization methods in this paper are as follows: the noise of normal distribution is added to select of focus;the number of focus is constrained;the multi-scale local ROI is cropped;the model can be trained using the policy gradient descent method of the reinforcement learning and the optimal cross entropy loss function. The experimental results show that the recognition rate is as high as 99.3%. Compared with the traditional local feature extraction method, the recognition rate of the dorsal hand vein is greatly improved.
作者
王一丁
赵晨爽
Wang Yiding;Zhao Chenshuang(North China University of Technology, Beijing 100144, China)
出处
《计算机测量与控制》
2019年第7期200-204,共5页
Computer Measurement &Control
基金
国家自然科学基金项目(61673021)
关键词
机器学习
视觉注意力机制
循环神经网络
强化学习
视觉图像处理
machine learning
attention mechanism
circulating neural network
reinforcement learning
visual image processing
