Research On Fingerprint Orientation Field Recognition

Nowadays fingerprint recognition technology has made a lot of progress,and has entered the stage of productization.There are a wide range of applications in the fields of electronic products,security,and criminal investigation.However,in the field of criminal investigation,in many cases,due to the complexity of the fingerprint extraction environment and the long-term environmental erosion,the quality of fingerprint extraction is very poor,and even cannot be identified by the naked eyes.This has caused great difficulties for the confirmation work of suspects..For such low quality fingerprints,it is often necessary to assist in denoising by extracting the orientation field before the enhancement and alignment of the fingerprint.This project is aimed for application.The company's request is to make an algorithm system that can extract a more accurate fingerprint orientation field to replace the fingerprint orientation field extraction part of the original system,which is also the direct source of this project.This project designed the fingerprint direction field extraction system and solved a series of problems in the system.Finally,the performance of the system extracting the fingerprint direction field was tested and verified.First,for the overall design of the system,this paper gives three possible solutions,namely direct decoding,clustering scheme and traditional visual + deep learning scheme.After verification,the direct decoding and clustering scheme were excluded due to poor performance on the latent fingerprints.Second,we designed the image preprocessing part of the system using traditional visual algorithms.The preprocessing is divided into two parts.The first is the filtering of the main noise in the image.Cartoon-texture decomposition effectively extracts texture parts from the image,but still retains a lot of texture noise.Most of the texture noise can then be filtered out using frequency domain filtering.The second part is the extraction of the effective area of the fingerprint.Latent fingerprint images contain a large number of non-fingerprint parts,and many areas of fingerprint texture damage are too serious,making it difficult to restore.Masking these areas has a significant improvement in the accuracy of fingerprint direction field prediction.Third,we also designed a deep learning network used for extracting the orientation field from the preprocessed fingerprint image.The origin of the network is a semantic segmentation network,and we made several optimizations to solve the problems met in this work.After the network output is changed to the regression type,the convergence speed and accuracy are greatly improved.After adding the machine learning boosting algorithm,the problem of the angular gap is solved.Finally,this topic also designed the corresponding loss function for this project.Finally in this paper,corresponding quantitative test experiments are also made.In the final algorithm of the project,the top-1 accuracy exceeds 0.85 even when using weak tags,which meets and exceeds the accuracy of the original system,which fully proves the generalization ability and reliability of the algorithm.