Bit-string representation of a fingerprint image by normalized local structures

•Due to the proposed normalized local structure, two central minutiae in the two local structures are now resided in the Euclidean space, and their dissimilarity can be computed by the Euclidean distance (ED) in pixel-based space.•Each minutia (reference minutia) in a fingerprint image is represented by a collection of neighboring minutiae enclosed within a circle, referred as a minutia-based local structure (MBLS).•A texture-based local structure (TBLS) is devised and fused with the MBLS via a weighted feature-level fusion scheme to improve the intermediate fingerprint representation discriminative before bit-string conversion.•A K-means algorithm is applied to transform the fused MBLSs and TBLSs into a fixed-length bit-string.•A finger-specific bit-training mechanism is proposed to further enhance the accuracy on fingerprint bit-strings. Conventional minutia-based fingerprint recognition requires a complicated geometric matching and hard to be adopted in the bit-string based cancellable biometrics or bio-encryption, as the minutia data representing a fingerprint image is geometrical, unordered and variable in size. In this paper, we propose a new method to represent a fingerprint image by an ordered and fixed-length bit-string to cope with those difficulties with providing a faster matching, compressibility and improved accuracy performance as well. Firstly, we devised a novel minutia-based local structure modeled by a mixture of 2D elliptical Gaussian functions to represent a minutia in the image pixel space. Then, each local structure was mapped to a point in a Euclidean space by normalizing the local structure by the number of minutiae in it. This simple yet crucial computation for converting the image space to the Euclidean-space enabled the fast dissimilarity computation of two local structures and all followed processes in our proposed method. A complementary texture-based local structure to the minutia-based local structure was also introduced, whereby both were compressed via principal component analysis and fused in the compressed Euclidean space. The fused local structures were then converted to a K-bit ordered string using the K-means clustering algorithm. This chain of computations with the sole use of Euclidean distance was vital for speedy and discriminative bit-string conversion. The accuracy was further improved by the finger-specific bit-training algorithm, in which two criteria were leveraged to select the useful bit positions for match