Optimizing CNNs for Contactless Palmprint Recognition

Abstract

Palm printing and palm vein recognition are the newer spheres within the already developed biometrics sector. Although many time-honored techniques have been offered and seen successful implementation over the last twenty years, the deep learning methods still lack all-round development in the palmprint and palm vein recognition. This research intends to study further the strength of deep learning on palmiets and palm vein recognition via 2D and 3D. We performed thorough examination of 17 known convolutional neural networks (CNNs) utilizing multiple databases, e.g., one 3D palmprint database, five 2D palmprint database and two palm vein databases. Our trials cover various network architectures, learning rates, and layer configurations, taking into account not only single mode data but also mixed mode data simultaneously. Results prove that CNNs of classic format show good recognition capabilities, and recent models with improved accuracy show even better achievements. One of the classic CNNs that stands out is Efficient Net. If the recognition accuracy is evaluated, this is the top performer. On the other hand, even though the classic forms of CNNs are fairly good in recognizing various types of tumors, their accuracy is still lower than the traditional methods. “Abstract” is a necessary section in a research paper. It may be constructed by gathering main points (summary) from each section of the research paper.