Curved Gabor Filters for Fingerprint Image Enhancement

1 Introduction and Related Work

• Curved Gabor filters adapt their shape to fingerprint flow, smoothing along bent ridges and valleys rather than using straight orthogonal filter axes.The authors also suggest possible use on other curved structures, including muscle fibers, cell filaments, and annual rings.Source: §1 · p. 2

1.1 Fingerprint Image Enhancement

• Low-quality fingerprints remain difficult to enhance because unreliable local orientation and ridge-frequency estimates can create artifacts and increase identification or verification errors.Enhancement also aims to clarify ridge and valley structure while reconnecting broken ridges, separating falsely conglutinated ridges, and preserving ridge endings and bifurcations.Source: §1.1 · pp. 2–3
• The paper combines two orientation-field estimators, estimates ridge frequency from curved regions, and defines curved Gabor filters from those regions for low-quality fingerprint enhancement.It reports improved matching performance over existing enhancement methods on FVC2004 images.Source: §1.1 · p. 3

2 Orientation Field Estimation

• The method averages line-sensor and gradient-based orientations when their angle differs by less than 15°; otherwise, it marks the pixel missing.It reconstructs inner gaps and extrapolates nearby outer orientations up to a 16-pixel radius.Source: §2 · p. 4
• The combined orientation field performed better in verification than either individual estimate across all 12 FVC2000–2004 databases.The authors interpret the lower equal error rates as an indicator that the combined field contains fewer estimation errors.Source: §2 · p. 4
• The fusion strategy treats the two orientation estimators as judges: incoherent or abstaining judgments are reconstructed or extrapolated rather than averaged.Incoherence is defined by an angle greater than threshold t.Source: §2 · p. 4

3 Ridge Frequency Estimation Using Curved Regions

• Unlike the x-signature method’s straight 32×16-pixel window, the proposed RF method uses curved regions and adds reliability scoring, conditional smoothing, extrema pairing, and inverse-median estimation.Straight windows may lack sufficient information in noise, and enlarging them in curved areas can mix neighboring ridges and valleys.Source: §3 · pp. 4–5

3.1 Curved Regions

• Each region is built from parallel curves initialized orthogonal to local orientation and traced along it; the illustrated profile averages interpolated values across 33 lines of 65 points.The construction stops when consecutive local orientations change beyond a threshold, indicating a presumed core point.Source: §3.1 · p. 5, Figure 2 · p. 5

3.2 Curvature estimation

• The regions also yield a pixel-wise curvature estimate: integrated orientation change along each central curve, with core and delta points visible in the curvature image.The estimate sums absolute orientation differences between the central point and the two curve endpoints.Source: §3.2 · p. 6, Figure 4 · p. 7

3.3 Ridge Frequency Estimation

• Ridge frequency is estimated by taking the reciprocal of the median distance between consecutive profile minima and between consecutive maxima.The profile averages interpolated gray values along curves, then detects local extrema.Source: §3.3 · p. 6
• Large pmaxmin values flag unreliable extrema; estimates above 1.5 may be smoothed and recomputed, subject to extrema-count and valid-frequency constraints.The study used a Gaussian kernel of size 7 and σ = 1.0, with at most three smoothing iterations.Source: §3.3 · p. 7

4.1 Definition

• Each curved region is mapped to an array of interpolated gray values, then filtered with an unrotated Gabor function to produce the enhanced pixel.The curved region contains 2p + 1 parallel lines and 2q + 1 points per line.Source: §4.1 · p. 9
• Locally adaptive normalization compensates for brightness differences after filtering.The experiments set the desired mean and standard deviation to 127.5 and 100, using neighboring pixels within a radius of 16.Source: §4.1 · p. 9

4.2 Parameter Choice

• Small Gaussian spreads or windows can make straight Gabor filtering ineffective, while large ones risk introducing noise and false structures in curved flow.The same trade-off applies to window size and to σx and σy.Source: §4.2 · p. 9
• Curved Gabor filters allow larger regions and higher σx and σy without spurious features; their remaining limitation is orientation- and ridge-frequency estimation accuracy.Figure 6 illustrates scar removal and separation of falsely conglutinated ridges using σx = 8.0, σy = 8.0 and 65 × 65-pixel curved regions.Source: §4.2 · p. 9, Figure 6 · p. 10
• Circular-window and elliptical-region variants achieved similar results on the FVC2004 databases.The tested ellipse had major axis 2q + 1 and minor axis 2p + 1.Source: §4.2 · p. 10

5.1 Test Setup

• The tests used BOZORTH3 with MINDTCT and VeriFinger 5.0 Grayscale, following the FVC protocol with 2,800 genuine and 4,950 impostor attempts per database.Equal error rates were calculated using the cited procedure.Source: §5.1 · p. 10

5.2 Verification tests

• With the same orientation and ridge-frequency estimates and identical parameters, the straight-window filter created artifacts and deleted a true minutia, unlike the curved-region comparison.The only stated difference between the two filters was the underlying window shape.Source: Figure 7 · p. 11
• The evaluation tested curved-region sizes, Gaussian spreads, and interpolation methods against traditional Gabor, STFT, and pyramid-based enhancement using BOZORTH3.Bilinear and bicubic interpolation did not noticeably improve on nearest-neighbor interpolation in the reported tests.Source: §5.2 · pp. 10–11
• With x-signature ridge-frequency estimation and the authors’ orientation field and segmentation, EERs were lower on DB1 and DB2, higher on DB3, and similar on DB4 versus the cited implementation.Compared with original images, performance improved on DB1 and deteriorated on DB3 and DB4; incorrect x-signature frequency estimates largely caused DB3’s increase.Source: §5.2 · p. 11
• To the authors’ knowledge, combining MINDTCT templates from curved-Gabor- and anisotropic-diffusion-enhanced images achieved the lowest EERs yet on FVC2004 with BOZORTH3.The results are listed in Table 1.Source: §5.2 · pp. 11–12, Table 1 · p. 12
• Curved-Gabor smoothing improved VeriFinger verification despite its non-disableable built-in enhancement of both original and enhanced images.The software’s enhancement procedure was undisclosed.Source: §5.2 · p. 12

• The method estimates ridge frequency using curved regions and enhances low-quality fingerprints with curved Gabor filters, improving matching performance versus existing enhancement methods.Source: §6 · p. 13
• On a GPU, curved-region ridge-frequency estimation took about 320 ms and applying curved Gabor filters took about 280 ms, indicating practicability for online verification.The tested region size was 33 × 65 pixels and the filter size was 65 × 33 pixels.Source: §6 · p. 13
• Further matching gains may depend on better orientation-field estimation; future work includes locally adapting filter parameters and applying the method to latent fingerprints.The authors note that filter size and Gaussian-envelope standard deviations require care to balance smoothing against spurious features while orientation-field errors remain.Source: §6 · p. 13