Research On Image Quality Evaluation Algorithm Of Finger Vein Based On Multi-label CNN

Finger vein biometrics has been widely used for personal authentication.Some low-quality images generated during acquisition may have a serious impact on subsequent feature extraction and matching.It is necessary to evaluate the quality of the acquired images.Traditional quality assessment methods are more difficult to select manually designed features,and there are also problems such as the complexity of the process of detecting venous feature points.More and more researchers have applied deep learning to finger vein quality assessment.At present,the related work of deep learning is aimed at two or more types of tasks,but in the actual acquisition process,affected by environmental lighting,finger posture changes,etc.,a low-quality image often contains one or more labels,Two-class or multi-class methods may not solve this problem.In addition,the existing quality evaluation methods all remove low-quality images and leave high-quality images for identification after performing the quality evaluation,and the image utilization rate is not high.Aiming at the above problems,this paper proposes a multi-label deep neural network for the quality evaluation of finger vein images.The specific research contents of this paper are as follows:1)Make a multi-label data setAiming at many low-quality factors such as rotation,translation,and noise that may exist in low-quality images during actual acquisition,we artificially rotate,translate,and change the brightness and noise of finger vein images to produce a data set of multi-labeled finger vein images.2)Improved Single-Scale Retinex image enhancement algorithmIn the traditional Single-Scale Retinex(SSR)algorithm,a Gaussian surround function is used to estimate the illumination of the image,but when there are a large number of concentrated gray areas in the image,the image enhanced by this algorithm will have a "halo" phenomenon.In this paper,weighted guided filtering with edge sensing weights is used instead of Gaussian filtering to overcome the above problems.3)Introducing the channel spatial attention moduleIt is used to weight the channel and spatial information of the output from each CNN block to enhance the representation capability of the feature map.By merging the results of maximum pooling and average pooling,richer features are obtained for multilabel classification.4)Convolutional neural network for multi-label quality evaluationThe single Softmax of the improved network output part is replaced with 4 Softmax loss functions for training on multi-label data sets.After that,the low-quality image after the quality evaluation is transformed into a high-quality image using the Bspline registration and fusion filtering image registration method for recognition,thereby improving image utilization.The experimental results show that the quality evaluation algorithm proposed in this paper has greatly improved the classification accuracy,and the accuracy and image utilization of vein authentication after registration are far better than some current quality evaluation algorithms.