OpenCV-Based Finger Vein Image Matching: Techniques and Workflow

First‑generation biometric methods and their limitations

Face, fingerprint, voice and palm‑print recognition rely on external physiological traits. These traits can be spoofed with forged images, videos or artificial models, so systems must add liveness detection to ensure the presented sample originates from a live user.

Finger‑vein recognition advantages

Finger‑vein systems illuminate the finger with near‑infrared light; hemoglobin absorbs the light and reveals the vein pattern beneath the skin. Because the pattern is internal, it is unaffected by surface conditions such as moisture, dirt or cuts, and it is extremely difficult to replicate, providing stronger anti‑counterfeiting security than first‑generation methods.

OpenCV image‑processing pipeline

3.1 Smoothing (Gaussian blur)

Gaussian blur replaces each pixel with a weighted average of its neighbours using a convolution kernel (e.g., a 5×5 mask). The kernel values follow a Gaussian distribution, giving the centre pixel the highest weight. The kernel size must be odd; sigmaX and sigmaY control the spread. Alternative smoothing filters mentioned are averaging, median (effective against salt‑and‑pepper noise) and bilateral filtering (preserves edges while reducing noise).

3.2 Histogram equalization

Histogram equalization redistributes pixel intensities so that the histogram becomes approximately uniform, enhancing contrast while preserving the relative ordering of pixel values. The mapping function must keep output values within the 0‑255 range and must not alter the ordering of brightness levels. The process uses a discrete cumulative distribution function: for each gray level k, the new value s_k = (L‑1)·(∑_{j=0}^{k} n_j)/n, where n_j is the pixel count for level j, n is the total number of pixels, and L is the number of possible gray levels.

3.3 Edge detection

Edges are locations with significant intensity changes. After smoothing, edges can be detected with first‑order operators (Sobel, Roberts, Prewitt, Kirsch) or second‑order operators (Canny, Marr‑Hildreth, Laplacian). The Sobel operator uses two 3×3 kernels to compute gradients in the x and y directions:

Gx = [[-1, 0, 1],
      [-2, 0, 2],
      [-1, 0, 1]]
Gy = [[-1,-2,-1],
      [ 0, 0, 0],
      [ 1, 2, 1]]

The gradient magnitude is obtained by combining Gx and Gy, and a threshold is applied to keep strong edges.

Feature detection and matching with OpenCV

4.1 Keypoint extraction

OpenCV represents image features as keypoints (position, scale, orientation) and descriptors (vectors describing the surrounding pixel pattern). SIFT (Scale‑Invariant Feature Transform) extracts keypoints through four steps:

Scale‑space extrema detection to locate potential keypoints.

Keypoint localization to filter unstable points.

Orientation assignment for rotation invariance.

Descriptor generation (128‑dimensional vector).

Alternative detectors mentioned are SURF, ORB and BRISK.

4.2 Descriptor matching (knnMatch)

The knnMatch function returns the k best matches for each query descriptor. Setting k=2 yields the closest and the second‑closest match; the ratio of their distances is compared against a predefined threshold (Lowe’s ratio test). Matches that satisfy the ratio are kept as reliable correspondences. In finger‑vein verification, the number of retained matches indicates whether two images belong to the same finger.

4.3 Complete matching pipeline

The full workflow consists of:

Pre‑processing: Gaussian blur → histogram equalization → edge detection (Sobel).

Keypoint extraction (e.g., SIFT).

Descriptor matching with knnMatch and ratio test.

Parameter choices—kernel size for blur, histogram mapping limits, Sobel gradient thresholds, and the Lowe ratio—directly affect the quantity and quality of extracted keypoints and thus the final matching outcome.

Examples show original, pre‑processed and matched images for the same finger (blue circles = keypoints, green circles + lines = matches) and for different fingers, demonstrating discriminative power.

References

陈华宣. 指静脉识别系统原理概述[J]. 电子技术与软件工程, 2016(22):81.

OpenCV Documentation: https://docs.opencv.org/4.0.0/index.html