Latent Fingerprint Enhancement Using Deep Learning

Latent fingerprints are fingerprints left at the scene of a crime without meaning to.Compared with conventional fingerprint,the image quality of latent fingerprint is poor,the texture is not clear and is often covered by various noises.Traditional automatic fingerprint identification system is difficult to deal with these characteristics.Therefore,the latent fingerprint enhancement and feature extraction methods with high performance have become the hot issues in this field.At present,the latent fingerprint enhancement algorithm based on the traditional method is not robust enough for the complex background of the image.Due to the unreasonable network design and training methods,the method based on deep learning still has much room for improvement.Based on the latest development of deep learning,this paper proposes a new method of latent fingerprint enhancement based on generating antagonistic network.As a kind of generation model,Generative Adversarial Network has achieved excellent performance in many Computer Vision tasks.The latent fingerprint enhancement model in this paper is mainly composed of two parts: a generation network composed of convolution layer and deconvolution layer and a discriminator network.The convolution part is the extraction of fingerprint features,especially for enhancement.Enhanced deconvolution technology is used to remove structural noise and enhance fingerprint.By using pixel-to-pixel and end-to-end learning,the enhanced network can directly enhance the latent fingerprint as output.We also studied some implementation details such as artificial latent fingerprints and multi-scale image fusion.Through experiments on NIST SD27 latent fingerprint datebase and the Verifinger SDK as the matching algorithm,the recognition rate of ran-1 and rank-20 reached 41.5% and 59.30% respectively,indicating the effectiveness and robustness of the method in the latent fingerprint enhancement.Fingerprint features play an important role in fingerprint identification.It is difficult to extract reliable details from poor quality fingerprints.At present,the mainstream minutiae extractation algorithm is aimed at the high quality fingerprint,which cannot satisfy the latent fingerprint minutiae extraction under the complex situation.Although there are some latent fingerprint minutiae extraction algorithms,due to the backward network architecture and the lack of domain knowledge,the accuracy of minutiae extraction cannot meet the requirements of system identification.Aiming at these problems,this paper designs and implements a method of latent fingerprint minutiae extraction based on domain knowledge and cascading deep network.It mainly includes two parts,the Minutiae Pre-Extraction Network based on and the Minutiae Discriminator Network based on Convolutional Neural n Network(CNN).By taking the scene fingerprint enhancement image as input,the influence of complex background noise can be overcome.By using domain knowledge,the training network can learn the features of minutiae directly from the fingerprint data.The original fingerprint is mapped at a fixed step size to the corresponding minutiae fraction.Therefore,a large number of minutiaes will be extracted by a given threshold.Then the minutiae blocks with these details as the center are input into the convolutional neural network(CNN)to rejudge and calculate the direction of these details.Experimental results show that the Precision Rate of the proposed algorithm is 66.7% and the Recall Rate is 68.2% for detail extraction on the public NIST SD27 field fingerprint database,which is significantly higher than other advanced minutiae extraction algorithms.