Research On Low Quality Fingerprint Image Processing And Feature Matching

A low-quality fingerprint is usually defined as a fingerprint characterized by vague, high-level noise or low contrast. Given that both false reject and false accept produced by non-linear distortion greatly degrade the performance of a fingerprint verification system, a fingerprint image that is clear but is with large deformation is also classified as a low-quality fingerprint in this dissertation. Although there have been great progress in studies on fingerprint verification, fingerprint verification under low quality conditions still have not been unable to meet the requirements of numerous practical applications. The specific difficulties include poor image processing results, as well as false reject and false accept led by non-linear distortion, etc. These limit the application of fingerprint verification. In such circumstances, further improvements of the performance of fingerprint identification have important significance both in theory and application. On the basis of the systematic analysis of previous research results, this dissertation proposes an in-depth study of low-quality fingerprint image processing and feature matching technology. The details are provided below.1. To address the shortcoming brought about by traditional fingerprint image segmentation algorithm having difficulty in segmenting low-quality fingerprint image and need-tuning parameters, according to the fingerprint texture being with parallel ridge and valley and the ridge orientation in a local region being basically the same, this dissertation proposes Fingerprint Texture Intensity feature and use this feature to segment fingerprint image. The algorithm first calculated the Fingerprint Texture Intensity according to the ridge orientation and the ridge circle. The adaptive segmentation threshold was then obtained according to the distributions of the Fingerprint Texture Intensity, and the image was segmented by this adaptive threshold. Finally, the segmentation noise in the foreground and the background were eliminated by a morphological algorithm. The tests conducted on the Fingerprint Verification Competition (FVC) databases showed that this algorithm performed well for low-quality fingerprint image. Moreover, neither training process nor manually changing parameters were needed for segmenting images from different kinds of sensors. 2. To address the shortcoming brought about by most fingerprint image enhancement algorithms either producing excellent enhanced results but with high computational time cost, or producing non-ideal enhancement results with low computational time cost, this dissertation proposes a fast fingerprint image enhancement algorithm that achieves a good trade off between enhancement result and computational time cost. In comparison with conventional enhancement algorithm, a two-level convolution mask was used in this algorithm instead of a multi-level convolution mask. The algorithm consisted of three procedures: (1) a two-level convolution mask was designed according to the ridge circle and the ridge orientation; (2) the enhancement results of the beginning pixels in each block were obtained by comparing two average gray-levels corresponding to the two parts of the convolution mask; and (3) except for the beginning pixels in each block, the enhancement result was obtained by modifying the enhancement result of the previous pixel which resulted in faster enhancement algorithm.3. To address the shortcoming brought about by most fingerprint features having low discrimination, and the corresponding feature extraction algorithm having high computational time cost, this dissertation investigates an Adjacent Feature Vector-based description method for fingerprint verification. We likewise propose a feature-matching method based on the Adjacent Feature Vector (AFV). This dissertation describes a fingerprint through the location, direction, and AFV of each minutia. An AFV consists of four adjacent relative orientations and six cut-ridge numbers. The AFV-based minutiae-matching method consisted of three procedures: (1) calculating the matching similarity level of any minutiae-pair and addition of the minutiae-pairs whose similarity levels were greater than a specified threshold to possible match minutiae-pairs set; (2) fetching each minutiae-pair in possible match minutiae-pairs set and performing coordinate translate and coordinate rotation to align the fingerprint, respectively; and (3) calculating the overall similarity of fingerprint matching by summing up the matching similarities of each minutia. The tests conducted on the FVC databases showed that AFV was a reliable feature and that the AFV-based matching algorithm performed well.4. To address the shortcoming brought about by the traditional fingerprint feature matching algorithm producing poor results for fingerprint with non-linear deformation, this dissertation investigates a Local Relative Location Error Descriptor-based method for overcoming non-linear deformation. The distortion is usually very small, and the relative position is basically the same in a local region when non-linear distortion occurs. Based on this property, this dissertation proposes a descriptor called the Local Relative Location Error Descriptor (LRLED), and proposes a feature-matching algorithm that can overcome non-linear distortion. The LRLED-based minutiae-matching method consisted of three procedures: (1) a pairwise alignment method was proposed to achieve fingerprint alignment; (2) a matched minutiae-pair set was obtained with a comparatively loose threshold to reduce false non-matches which led not only to most of the corresponding minutiae-pairs but also to a few non-corresponding minutiae-pairs being matched; and (3) an LRLED-based similarity measure was employed to compute the similarity level between template and test fingerprints which produced a very high score for a corresponding minutiae-pair but a very low score for a non-corresponding minutiae-pair to reduce false matches. Evaluations on the LRLED feature and the matching algorithm, as well as some matching examples all demonstrated that the LRLED-based minutiae matching can overcome non-linear distortion. The matching results can be considered promising.The above researches improve the performance of fingerprint image processing and feature matching, and are helpful for expanding the application field of fingerprint verification. In addition, the proposed method for overcoming non-linear distortion provided a new idea for a point-matching problem under the condition of non-linear distortion.