Research On Contactless Palmprint Recognition Based On Learnable Gabor Convolution

With the development of biometric recognition technology,palmprint recognition has attracted extensive attention from domestic and foreign researchers due to its rich features and advantages of convenience and safety.Research on contact-based palmprint recognition has been fruitful,in comparison,the research on contactless palmprint recognition with low cost,high freedom and easy expansion has encountered some challenges.Compared with traditional contact-based palmprint recognition methods,contactless palmprint images collected in free space are often accompanied by a certain degree of rotations,noises,lighting and complex background.However,none of the existing palmprint recognition algorithms can perfectly deal with these problems,which limits the accuracy of contactless palmprint recognition.Therefore,this article has conducted extensive investigations on the shortcomings of current contactless palmprint recognition algorithms,and conducted in-depth research on palmprint image rotation estimation,key point detection,and feature extraction.Mainly completed the following work:Based on rotation estimation,a palmprint image Region of Interest(ROI)extraction algorithm id designed.The algorithm estimates and corrects the rotation angle of the contactless palmprint image through a convolutional neural network,thereby overcomes the serious problem of contactless palmprint image rotation,which greatly improves the reliability and accuracy of palmprint image ROI extraction.Experimental results show that the algorithm accurately and effectively corrects the rotation of the palmprint image,and achieves better results than the classic palmprint ROI extraction algorithm.A palmprint image ROI extraction algorithm based on small targets segmentation is proposed.The algorithm regards the key points detection of palmprint image as small targets segmentation.It predicts the location of the key points directly from the pixel level,and uses a multi-branch network to learn features of different scales.By performing multi-resolution fusion,it ensures the final positioning accuracy.Experimental results show that this method effectively overcomes the problem of positioning key points of contactless palmprint images,and ensures the robustness of the algorithm in complex environments and complex poses.A feature extraction method based on a learnable Gabor convolutional neural network is proposed,and a series of Gabor convolution kernels are learned to deeply mine the rich direction information of palmprint images.At the same time,the algorithm uses the Angular Margin Loss to constrain the feature vector to keep it highly distinguishable in the angle space,which promotes the further improvement of recognition accuracy and ensures the compatibility of the algorithm with the open set.In addition,the algorithm uses convolutional downsampling instead of statistical pooling operators to ensure the stability of the direction information before and after pooling.The results of comparative experiments in this paper show that the algorithm has higher recognition accuracy than classic algorithms or algorithms in recent years under most public datasets,which illustrates the superiority of this method.