Research On Fingerprint Singular Points Detection Algorithm

Automatic fingerprint identification technique has become the most mature biometric identification technique because of its long history and high identification accuracy. In large fingerprint database, it is time consuming that input fingerprint compares with the reference fingerprints in database. Fingerprint classification is the key technology to speed up fingerprint identification algorithm, and it provides an indexing mechanism. Nowadays, most classification techniques are realized based on the information of singular points number, type, and location. In addition, in the process of fingerprint matching, algorithms based on singularity and texture require to extract singular points accurately and reliably. Therefore, accurate and reliable detection of singular points is necessary to fingerprint classification, even to the entire fingerprint identification system. Domestic and foreign scholars have done much study on singularity detection, but it is still not ideal in low quality fingerprint. To deal with the difficulty extracting singular points from low quality fingerprint image, three algorithms are presented from three different views in this paper.The main content of this article contains three parts, the improved Poincare index, singularity extraction combined with Gaussian-Hermite moment distribution, singularity detection based on multiresolution.Improved Poincare index: to heighten antinoise capability of traditional Poincare index and reduce false singularities, an improved algorithm for Poincare index is proposed. As directions of the points on the closed curve around singularity are continuous, and there is only one absolute value of difference between adjacent angles is greater thanπ/2, the formula of adjacent direction angles difference is adjusted. To some extent, this enhances anti-noise capability of traditional Poincare index and decreases the number of undetected singularities.Singularity extraction combined with Gaussian-Hermite moment distribution attribution: in dealing with low quality fingerprint image, as extracting reliable direction field itself is a difficulty problem, singularity detection algorithms simply based on direction field information not only weaken antinoise capability but also detect many false singular points. Gaussian-Hermite moment is an orthogonal moment with smoothing window function, so can better describe ridge coherence information in low quality fingerprint image. The ridge coherence in singular area has following characteristics: the farther the distance from the singularity is, the better the ridge coherence is, and vice versa. Based on this global information, algorithm can delete false singularity effectively. Because this method effectively assembles information of ridge orientation and coherence in singularity's neighborhood, it can extract singularities in a comparatively accurate and reliable way. Experimental results show its effectiveness and robusticity. 500 fingerprint images from the NIST-4 database are used for an experimental test, and the accuracy rate on identifying singular points is 93.05% (96.93% for core points and 86.43% for delta points).Singular points detection based on mutiresolution analysis: Far away from the singular area, the ridge curvature is becoming more and more smaller, so fingerprint is a texture image with multiresolution characteristics. Firstly, singularities are detected on block images through shifting position of the first block image time after time at the same block size and the concentrative regions of singularities detected under different block positions are got and centroids of the regions are computed to gain the accurate positions of singularities. Then, reliability of the singularities detected above is determined with multilevel block sizes. The algorithm extracts singular points at different resolutions with three block sizes, block shift at the same size and improved Poincare index. In this method, the characteristic of the relative concentration of positions of the singularities detected through block shift and of the corresponding relationship of the singularities detected with multilevel block sizes are used to detect singularity accurately and reliably. Local direction field information is tap fully through block shift at the same block size, and multiresolution information is extract through different block sizes. This combination between local information and global information enhances antinoise capability of our algorithm. Experiment results on some typical low quality fingerprint images verify the effectiveness of the method.