Research On Local Invariant Feature Based Palmprint Recognition Method

As an emerging biometric recognition technology,palmprint recognition has received wide attention in the academia and industry over the past decade.The traditional contact palmprint acquisition device captures palmprint images under the pure background and the professional guidance,which guarantees high quality palmprint images at the expense of safety,hygiene and user-friendliness.Contactless palmprint recognition is efficient and user-friendly,but the low constraint of contactless acquisition device causes palmprint deformation.In addition,the mainstream region of interest(ROI)extraction algorithm can effectively extract ROI from the collected palm images,but it still has some limitations.Aiming at the too many preset parameters problem in ROI extraction and the deformation problem in palmprint images,this dissertation proposes a palmprint recognition method based on local invariant features to improve the accuracy of the palmprint recognition system.The main works are as follows:1)The palmprint recognition based on local invariant features is reviewed: its characteristics in three stages: local invariant feature detection,local invariant feature description and local invariant feature matching are analyzed.2)Region of Interest extraction of palmprint image: Aiming at the problem that there are too many preset parameters in ROI extraction,an effective ROI extraction algorithm based on linear clusters is proposed.The original whole hand image is first converted to a binary image.In binary image,we draw many straight lines according to several predefined rules.Then,we draw many straight lines to detect finger joint areas,which may result in detecting several different key point candidates in one finger joint area.Thus,a lot of key point candidates may be obtained in four finger joint areas.We then exploit K-means clustering algorithm to calculate four cluster centers,which are treated as the final four key points.The final key points can be used to construct coordinate system.In this new coordinate system,after rotation normalization,the ROI can be extracted from the central region of hand.We also collected a database including 16000 whole hand images,and experimental results indicate that the proposed method can achieve 100% localization and extraction accuracy3)For the deformation problem of contactless palmprint images,the palmprint recognition based on SIFTGPU and vector field consensus algorithm is proposed.First,a circular Gabor filter is applied to enhance the palmprint images.Then,the palmprint images feature extraction and matching are performed by using the scale-invariant feature transform(SIFT)which maintains the invariance of image scaling,rotation and even affine transformation,and GPU used to accelerate this process.Finally,a vector field consensus method(VFC)is proposed to refine the mis-matched points.The experimental results on the public palmprint datasets show that the proposed method can effectively improve the accuracy and speed of contactless palmprint recognition.