Abstract: Based on analyzing and experimenting with a large number of existing fingerprint image preprocessing algorithms, this paper presents a novel high-quality fingerprint image preprocessing algorithm. Taking full account of the inherent characteristics of fingerprint images, the algorithm balances overall features and local detail. Experiments demonstrate that it is a simple, efficient algorithm which improves the quality of fingerprint images and lays a solid foundation for subsequent processing.
Introduction
Automated Fingerprint Identification Systems (AFIS) rely heavily on the quality of input images. Raw scans are frequently plagued by noise, low contrast, and non-essential background data. Traditional preprocessing methods often suffer from detail loss or the creation of artifacts like "burrs" during the thinning process. This paper proposes a multi-stage pipeline designed to be computationally efficient and robust against real-world noise.
Methodology
The proposed algorithm consists of four distinct stages designed to transform a noisy grayscale scan into a clean, single-pixel-width skeleton.
1. Background Extraction
The image is divided into 20x20 blocks. By calculating the variance and contrast for each block, the system adaptively distinguishes between high-variance ridge areas and low-variance background noise, allowing the processor to ignore non-essential regions.
2. Directional Smoothing
To avoid blurring ridge boundaries, the authors utilize a "bar template" and "square plate" method. This approach targets noise while preserving the directional flow of the fingerprint ridges, ensuring that structural integrity is maintained.
3. Dynamic Binarization with Hole Padding
Instead of a global threshold, the system uses a 13x13 sliding window to calculate local thresholds. A critical "hole padding" step is included: if a 3x3 neighborhood is inconsistent (e.g., a white center in a black region), it is filled to ensure ridge continuity and prevent fragmentation.
4. LUT-Based Thinning
To reduce ridges to a one-pixel width, the algorithm employs an 8-neighborhood Look-Up Table (LUT). For any pixel $P0$, the system examines its eight neighbors ($P1$ through $P8$).
A unique index $S$ is calculated based on these neighbors using the formula $S = \sum_{i=1}^{8} Pi \times 2^{i-1}$. This index is used to instantly query a pre-defined array to decide whether to delete or keep the pixel. This space-time tradeoff significantly accelerates the thinning process compared to iterative logical calculations.
Key Results and Discussion
The algorithm was tested on 320x320 pixel fingerprint images using MATLAB. The results show a marked improvement in image clarity and structural consistency.
As shown in the processed results, the skeletonized image is remarkably clean. The combination of local adaptive binarization and the LUT-based thinning effectively eliminates "burrs"—tiny, jagged branches that often lead to false minutiae detection in standard algorithms. By ensuring ridge continuity through hole padding, the final output provides a high-fidelity map for feature extraction.
Conclusion
The proposed preprocessing pipeline offers a robust solution for fingerprint image enhancement. By prioritizing local context over global thresholds and utilizing efficient data structures like Look-Up Tables, the algorithm achieves high processing speeds without sacrificing detail. This method provides a reliable foundation for the subsequent stages of automated recognition and matching.
Summary updated