Research On Fingerprint Segmentation Appraisal And Fingerprint Segmentation With Semi-Supervised Learning

Automatic Fingerprint Identification System (AFIS) can quickly recognize a person by collecting, analyzing and matching fingerprints with special transfer equipment and image processing technology. Fingerprint segmentation is one of the key technologies in the pre-processing step of Automatic Fingerprint Identification System. Among current researches there are a variety of segmentation algorithms, however, there is no systematic application appraisal system. Machine Learning is one of the hot topics in Artificial Intelligence, and it is an application-promoted discipline. Current researches show that, many application problems can be solved by machine learning methods which are taken as quite effective instruments. One of these application problems is fingerprint image segmentation. This thesis proposes several researches on machine learning application problems in fingerprint segmentation.This thesis proposes a "Three Character—Double Level" fingerprint segmentation application appraisal system. Here, Three Characters are defined as "Pixel", "Block" and "Overall"; Double Levels are defined as "Character Level" and "Method Level" We describe these six dimensions, summarize current algorithms and briefly analyze segmentation error rate and time complexity. By presenting significant characteristics of different methods, we find that current fingerprint segmentation has well-defined characters and well-operated algorithms which already have high accuracy and perform speed, while the applicability is to some extent lacked. This thesis applies semi-supervised learning in fingerprint segmentation, and proposes two fingerprint segmentation algorithms—CoSeg and TriSeg, based on co-training style. The two algorithms do training processes under the CMV system with two or three out of LabelBox, Support Vector Machine, LS- Support Vector Machine, which effectively utilize labeled data and unlabeled data. The experimental results show that both CoSeg and TriSeg can achieve a better segmentation under the insufficient information circumstance. In this thesis, we find that the future work can have following aspects:(1) Personalized character definition for low quality fingerprint image should be reached, and when applied to multiple fingerprint databases the applicability of character and algorithm should be improved. (2) When several sensor equipments work together the sensor interoperability problems should be well solved. (3) Algorithms based on two independent views need to be researched, such as frequency view and space view, or pixel view and block view. (4) Labeled or unlabeled data need to be selected automatically instead of using sampling methods.