Research On Multi-Type Sensors Oriented Fingerprint Image Segmentation

Due to the uniqueness, stability and convenience, fingerprint recognition has become a mainstream means of biometrics and has been widely used in many fields. Fingerprint identification includes fingerprint preprocessing, feature extraction, fingerprint matching and so on. Fingerprint image segmentation is an important preprocessing step, its aim is to separate the fingerprint foreground area from the background area. Effective segmentation does not only improve the accuracy of minutiae detection but also reduce the time of subsequent processing.Many kinds of fingerprint sensors have been applied in fingerprint identification system. Fingerprint images captured by different sensors are usually different in gray level, resolution, quality and image size. Most existing methods are designed to segment fingerprint images collected by a certain sensor. The performance of the methods significantly degrades when segmenting images captured by other sensors. There are mainly two reasons:firstly, the images captured by different sensors have different distributions in various feature space, the features can not well discriminate the foreground and background. Secondly, the model trained on one database collected by a certain sensor is not suitable to images from other sensors. These problems not only decrease the performance of the system, but also limit the application of the fingerprint identification in internet environment. This paper studies the above problems in depth, the contents are as follows:(1) Fingerprint segmentation research using the difference of sensors. This topic studies the relation between image and feature, proposes the adaptive feature selection problem in fingerprint segmentation. An adaptive feature selection method based on decision tree is proposed to select suitable segmentation features for images captured by different sensors. The solution to this problem not only enhances the adaptability to different sensors of segmentation method and improves the performance, but also saves feature extraction time and image segmentation time.(2) Fingerprint segmentation research eliminating the difference of sensors. This topic analyzes the sensor interoperability problem of segmentation features and proposes a two level feature evaluation method. The method includes the first level evaluation based on segmentation error rate and the second level evaluation based on decision tree. Features with good sensor interoperability are selected based on the evaluation results and segmentation performance is improved.(3) Fingerprint segmentation research based on single image. This topic presents a method to improve the automatic labeling based linear neighborhood propagation segmentation method. This method uses a more robust automatic labeling mechanic to provide reliable labeled samples, which not only improves the segmentation performance but also better deals with sensor interoperability problem.The future work includes:firstly, define the features with good discriminating ability, provide a wealth of candidate features for feature selection and feature evaluation method; secondly, combine the feature research with method research to find the segmentation method which can automatically evaluate and select features.