为减少手指静脉识别时间, 提出一种双重降维方向梯度直方图特征( Histogram of Oriented Gradient, HOG)结合支持向量机( Support Vector Machine, SVM)分类的手指静脉识别方法。针对传统 HOG算法特征维数高的问题, 首先通过 Fisher准则衡量梯度方向区间 HOG特征的分类能力, 然后使用序列前向选择法挑选出分类能力较优异的梯度方向区间构建部分方向区间 HOG特征, 最后使用主成分分析(Principal Component Analysis, PCA)降维。在公开的手指静脉数据库 FV-USM和 THU-FV上使用 SVM多分类器进行分类识别, 实验结果表明: 双重降维 HOG方法相较于 HOG＋PCA方法提取的特征维数降低了 40%, 识别时间减少了 29.85%, 识别准确率分别为 99.17%和 100%, 等误率分别为 1.07%和 0.01%。

为了提取指关节纹图像的纹理特征并进一步提高其识别精度, 提出一种基于改进Gabor小波和Tetrolet的指关节纹识别方法。首先, 利用某点邻域像素的融合幅值特征代表该点的Gabor幅值特征, 增强每个像素点之间的局部关联性; 其次, 将滤波后的指关节纹特征图像经过具有高稀疏性的Tetrolet变换以获取图像的最优稀疏表示; 最后, 利用带限相位相关(Band-Limited Phase-Only Correlation, BLPOC)算法提取指关节纹图像的互功率谱进行匹配。在PolyU-FK、PolyU-CFK、IITD图库得到的识别准确率分别为99.1300%, 98.8324%, 98.7937%, 最低等误率为1.4601%, 最短识别时间为0043 2 s。本文方法能够提高识别系统的性能, 具有可行性和有效性。

针对指静脉提取困难和识别精度不够高等问题,提出了一种基于ResNet改进的手指静脉识别方法。首先,使用深度超参数化卷积(DO-Conv)代替网络中的传统卷积,在减少模型参数的同时提高了网络识别率。然后,将空间注意力模型(SAM)和挤压激励块(SE-Block)融合,应用于改进的残差网络(ResNet)提取图像在通道和空间域上的细节特征。最后,使用标签平滑的交叉熵(LSCE)损失函数来训练模型,实现自动校准网络防止分类出现误差。实验结果表明,改进后的模型不易受到图片质量的影响,在公开数据库FV-USM和SDUMLA上的识别精度分别达到99.4919%和99.4485%,较之前的网络在精度上有明显提高。

针对目前指关节纹识别方法鲁棒性差的问题,提出了一种基于非下采样的Shearlet变换(NSST)和Tetrolet能量特征的指关节纹识别方法。首先,采用直方图均衡化调整图像的灰度,以减少光照分布不均对识别系统产生的影响。其次,利用NSST及其逆变换得到去噪后的重构图像,并对其进行Tetrolet变换,建立低频图像的能量曲面。最后,将不同图像的能量曲面作差,得到能量差曲面,进一步计算曲面的方差,并以此为依据对不同指关节纹图像进行分类识别。在HKPU-FKP、IIT Delhi-FK、和HKPU-CFK图库及其噪声图库的实验结果表明,本方法的正确识别率可达98.0392%,最短识别时间为0.0497 s,最低等误率为2.5646%。相比其他方法,本方法可以明显提高指关节纹识别系统的性能,具有可行性和有效性。

Creatinine (Cr) is a biochemical waste molecule generated from muscle metabolism and primarily cleared from the bloodstream by the kidneys. If kidney function declines, Cr levels in the blood tend to increase. Therefore, Cr serves as an indicator of kidney function. In this work, we present a simple method for the rapid screening for impaired renal function based on the subject's Cr concentration. In our setup, broadband white light is delivered to a finger clamp through a fiberoptic cable to illuminate the patient's finger. The light is transmitted through the finger and collected by a second optical fiber coupled to a visible–near-infrared (VisNIR) spectrometer which covers the spectral range from 400 nm to 1100 nm. During the calibration process, the transmitted spectra acquired from 60 patients were measured. An average was calculated using the peak level of the transmitted, diffused intensity at three different wavelengths to create a “Cr intensity index". Patients were divided into five groups according to their Cr concentration levels, ranging from 1 mg/dL to 13 mg/dL. Our observations indicated that each group featured a unique spectral fingerprint. Next, we tested the index on 20 patients not included in the calibration procedure (unknown samples). We were able to classify patients into groups according to their Cr level with moderate prediction accuracy (R² = 0.55) and mean screening error of up to 16%. Future efforts will evaluate the accuracy of this approach with larger patient populations representing a broad range of Cr concentration. Still, this preliminary work is an essential step toward developing this useful noninvasive Cr screening platform using NIR light spectroscopy.

We present a systematical study on comparison between water and dry coupling in photoacoustic tomography of the human finger joints. Compared to the direct water immersion of the finger for water coupling, the dry coupling is realized through a transparent PDMS film-based water bag, which ensures water-free contact with the skin. The results obtained suggest that the dry coupling provides image quality comparable to that by water coupling while eliminating the wrinkling of the finger joint caused by the water immersion. In addition, the dry coupling offers more stable hemodynamic images than the water coupling as the water immersion of the finger joint causes reduction in blood vessel size.