1. Summary

This repository consists of 3 modules:

A fingerprint pre-processing chain based on image filtering algorithms
An algorithm for extracting fingerprint’s features, which are called “minutiae”
A one-to-one fingerprint matching algorithm.

2. Dataset

The dataset used in this repository consists of 500 fingerprint images, each has the resolution of $320 \times 480$ pixels (Width $\times$ Height), white ridges & black background. You can see the dataset here.

3. Pre-processing module

The first module of this repository transforms a fingerprint image into its skeleton, using Gabor Filter and Morphology Skeletonization. Here are some of its results:

Pipeline

The implemented pipeline is based on this article. The following images show the steps in order, as well as the result of each step:

Performance Status

In order to make the benchmark result accurate and reliable, the program is single-thread. The fingerprints will be processed one after another. There is no paralleled process.

Processed 500 images: 56.599s
Average time per image: 0.1132s
Images per second: 8.83 img/s

Development Status

The Gabor Filter is not fully optimized. In fact, 80.35% of the runtime is spent on a single step in the pipeline.

Documentation for this module will be composed later.

4. Feature Extraction Module

This module proposes a simple feature extraction algorithm. It takes a skeleton image of a fingerprint, then returns the terminations and bifurcations. It also comes with a little noise elimination, which increases the accuracy of the matching module.

5. Matching Module

Algorithm: Improving Fingerprint Verification Using Minutiae Triplets.

Benchmark Proposal

To read more about the matching benchmark, see _doc/

This benchmark uses the same dataset as the one in the Pre-processing section. The program is then challenged:

Given $N$ fingerprints. For each fingerprint, find another one that is of a same finger.
Because the project aims at fingerprint recognition, not fingerprint grouping, each comparison must be one-to-one. Thus, for this dataset, there are $500 \times 499$ comparisons expected.

Again, the program is single-thread, which ensures the average time for each comparison is accurate and reliable. Each comparison is performed one after another. There is no paralleled process.

Benchmark - Accuracy

481 out of 500 given fingerprints successfully found an accurate match. That makes the accuracy of 96.2%, or 3.8% mismatch rate. Here are some of the matches:

Benchmark - Performance

Executed $500 \times 499 = 249500$ comparisons: $37.738$ s
Average $0.1512 \times 10^{-3}$ s, or $0.1512$ ms per comparison
Speed: $6611.37$ comparisons/s

Which sucks, tbh.

8. How to use

Because I haven't made a cli yet, just edit Program.cs for the time being.

For example:

// create a new Util.PreprocessingMultithread.Processor object
Processor proc = new();
// load a fingerprint image
Image<Gray, byte> src = new("fingerprint-image.png");
// process it
proc.Process(src);
// export the fingerprint's skeleton
CvInvoke.Imwrite("skeleton.png", MatConverter.Bool2Img(proc.SkeletonMat));
// export the fingerprint's minutiae
proc.PrepareForExtraction();
MinutiaeExtractor.ExtractAndExport("data-1.inp", proc.SkeletonMat, proc.SegmentMsk);

// load fingerprint's data & match them
Fingerprint probe = new("data-1.inp");
Fingerprint candidate = new("data-2.inp");
Console.WriteLine(Matcher.Match(probe, candidate)); // can be visualized via Matcher.DebugMatch(probe, candidate, probeMatchResult, candidateMatchResult)

7. Directories

If you're really into the source code, here's the hierarchy:

./Algorithm: Algorithms for general purposes.
./DataStructure: Data Structures for general purposes.
./MatTool: Classes that handle operations on matrices. Providing 2D Forward Loops, mathematic operators on matrices, matrix converters and more.
./Util: Classes that are specialized in processing Fingerprint and Fingerprint only. Currently includes:
- ./Util/Comparator: Compare fingerprints
- ./Util/FeaturesExtractor: Extract minutiae from skeleton images
- ./Util/Preprocessing: Pre-process the input images into ridges' skeleton
./_doc: Document files

Regardless, this project's implementations and the algorithms behind them are absolutely incomprehensible to anyone without documentation (except me, ofc). Therefore, I'll try to write as many docs as I can to explain my code, I promise.

Juweriomohamed/FingerprintRecognitionV2