To address the erroneous and insufficient extraction of minutiae information in fingerprint matching,this paper proposes an improved end-to-end multiscale inverted residual network for fingerprint minutiae extraction(IRFingerNet)based on deep learning.This improved residual structure is incorporated into the model,and lightweight networks that are easily optimized is built to reduce information loss while increasing the network depth.The multiple features of fingerprint are sent into the network as united features to enhance semantic information and improve the perception of details.The channel attention mechanism is applied to calibrate the united features so that the effective feature weight is increased and the invalid feature weight decreased.Experimental results on NIST 4,FVC 2002 and FVC 2004 databases show that IRFingerNet can perform the task of fingerprint detail point extraction more effectively in practical applications with higher accuracy and callback rate,and the overall F1 score strikes 0.87,making an efficiency 11%greater than the traditional extraction method,and a detection speed of 0.23 s per fingerprint image.
NETL
NSTL
万方数据
