Research On Key Technologies Of Low Quality Fingerprint Recognition

With the rapid development of information society,personal identification has higher requirements on the accuracy,security and practicality.Fingerprint recognition technology is currently the world's most widely used,the highest market share of identification methods because of its easy to collect,accurate identification,and high degree of public acceptance.So far,many researchers and institutions have done in-depth studies on the fingerprint recognition technologies,fingerprint recognition performance has been greatly improved.However,for low-quality fingerprint images,such as the latent fingerprint image collected by the criminal investigation site,it is still difficult to meet the practical requirements because of the large noise disturbance,serious deformation and incomplete fingerprint structure.Automatic fingerprint recognition includes image acquisition,image processing,feature extraction and feature matching.This thesis focuses on two key technologies in low quality fingerprint recognition:fingerprint segmentation and fingerprint direction field estimation.The main work achievements and innovations are as follows:1.Latent fingerprint segmentation based on linear density.Fingerprint image segmentation is an important processing step for more accurate and efficient feature extraction and fingerprint recognition by separating the useful fingerprint texture area from the complex background noise area.The useful fingerprint information is composed of the ridge and valley lines with a certain flow potential.According to the characteristics of the ridge line in the foreground fingerprint region in the low quality fingerprint images,this thesis proposes a fingerprint segmentation algorithm based on linear density.Firstly,the total variation(TV)model is used to decompose the fingerprint image into cartoon and texture parts,and then the line segments are located in the texture image and the line density map is calculated to express the ridge and valley structure information.Finally,a segmentation mask is generated by thresholding the linear density map.The experimental results on NIST SD27 latent fingerprint database show that the proposed segmentation algorithm based on linear density can effectively segment the useful fingerprint region from the background and improve the accuracy of fingerprint segmentation.2.Latent fingerprint segmentation by combining ridge density and orientation consistency.Compared to the background of latent fingerprint image,fingerprint foreground is characterized as concentrated ridge and orientation consistency.Based on the above observation,a fingerprint segmentation method is proposed to combine the ridge density and orientation consistency.First,a texture image is generated by decomposition of latent image with a Local Total Variation(LTV)model.And then,the ridge lines are detected and the ridge density and orientation consistency are computed and combined to generate a segmentation mask by thresholding.The fingerprint foreground region is segmented as the region of high ridge density and orientation consistency.Experimental results on NIST SD27 show that this algorithm can improve the segmentation accuracy when compared with the fingerprint segmentation method based on linear density.3.Studies on the orientation field estimation model based on discrete cosine transform(DCT)basis functions.Fingerprint orientation field is the global features of the fingerprint image,describing structure and direction information of the ridge and valley flow.It plays an important role on fingerprint enhancement,feature extraction and matching.An orientation reconstruction method is proposed based on sparse coding and Discrete Cosine Transform(DCT)basis functions.The proposed orientation reconstruction model does not need any prior information such as the locations of singular points and it is efficient to implement.More importantly,the effect of noise can be significantly reduced by sparse coding.In addition,we also proposed an orientation reconstruction method based on weighted DCT,which can suppress the noise better while keeping the orientation details of the singularity region.First,the DCT basis functions are used to build the basis atoms for linear representation of orientation field.Then,the DCT basis atoms of low and high orders are combined with the weights determined by singularity measurements for orientation reconstruction.The weighted DCT model is further extended for partial fingerprints to gradually and iteratively reconstruct the orientations in noisy or missing parts of fingerprints.The proposed method can perform well in smoothing out the noise while maintaining the orientation details in singular regions.The experimental results on NIST and FVC fingerprint databases show that the algorithm can effectively eliminate the noise while preserving the orientation details of the singular regions,and has good performance on the orientation reconstruction of low quality fingerprints.4.Studies on the orientation field estimation based on multi-scale dictionary learning and sparse coding.It is a learning based method including online learning stage and offline testing stage.Multi-scale dictionary learning is used in the offline stage,while sparse coding is used to estimate orientation in the online stage.The multi-scale dictionaries are built in the offline learning stage,which use the multi-scale patches extracted from high quality fingerprint images.While in the online stage,the gradient method is used to initialize orientation field,the multi-scale sparse coding is iteratively applied to reconstruct the orientation fields with gradually increasing the patch size.The small patch is used to maintain the orientation details of singular regions,while the large patch is used to suppress noise and restore the corrupted orientations.Experimental results on NIST SD27 latent fingerprint database show that the proposed algorithm based on multi-scale dictionary learning and sparse coding can effectively estimate the orientation field of low quality fingerprint and improve the accuracy of fingerprint recognition.