基于多图融合和改进Xception网络的跨设备手背静脉识别研究

Cross-Device Recognition Research of Dorsal Hand Vein Images Based on Channel Merging and Improved Xception Network

摘要:

手背静脉是一种新兴的生物特征识别技术,相比其他生物特征具有唯一性、防伪造性、稳定性、和非接触性等明显优势。由于采集设备和采集环境的不同,手背静脉灰度图像存在亮度、角度旋转、尺度缩放等差异,识别率较低。由此提出一种基于多图融合和Xception网络的手背静脉识别算法。首先在图像预处理后分割得到二值纹理图,然后将二值图转换为距离图,再由二值图细化得到骨架图。最后融合二值图、距离图和骨架图,得到包含纹理特征和形状特征的三通道合并图。采用Xception结构作为分类网络,并将其激活函数ReLU改为非线性更强的h-swish激活函数。相关实验在由实验室自建的1库和2库两个数据库上进行,其中1库作为训练集,2库作为测试集,最高识别率达到93.54%.

Abstract:

Recognition of dorsal hand vein is an emerging biometric identification technology, which has obvious advantages compared with other biometrics, such as uniqueness, anti-counterfeiting, stability, and non-contact. Due to the difference of the acquisition equipment and acquisition environment, the gray-scale images of the dorsal hand vein have differences in brightness, angle rotation, scale scaling, etc., so recognition rate is low. Therefore, a dorsal hand vein recognition algorithm based on multi-image fusion and Xception network is proposed. Firstly, a binary texture map is obtained by segmentation after image preprocessing, and then the binary image is transformed into a distance map, and then the skeleton image is achieved through thinning of the binary image. Finally, the binary image, distance image, and skeleton image are combined to obtain a three-channel merged image containing texture features and shape features. The Xception architecture is used as the classification network, and its activation function ReLU is changed to the more nonlinear activation function h-swish. Relevant experiments are carried out on two databases, library 1 and library 2, built by our laboratory. Library 1 is used as training set and library 2 is used as test set. The recognition rate reaches 93.54%.