TMS016_Spatial_Statistics_and_Image_Analysis

Assignments 2025

Project 3 Fingerprint-Siamese (FVC2000) 🖐️🔍

A minimal solution completed in a short time for fingerprint-matching on the public FVC 2000 / DB1_B subset. The model is a Siamese CNN trained only on DB1_B; given any two DB1_B .tif images, it decides whether they originate from the same finger.

Note on cross-database generalisation
We experimented with fine-tuning / zero-shot transfer to DB2_B or DB3_B, but the results were unsatisfactory. Consequently, the current code base does not include cross-subset migration utilities.

data/ and checkpoints/ are intentionally git-ignored.

Stage	Script	Main outputs
0️⃣ Preprocessing	preprocess.py	• Enhanced .tif (for inspection) • Normalised .npy (network input)
1️⃣ Pair generation	create_train_pairs.py	Train / val pair lists
2️⃣ Training	train.py	Model checkpoints (checkpoints/best.pt)
3️⃣ Validation	validate.py	ROC / metrics vs thresholds /
4️⃣ Batch inference	inference_batch.py	Match-score CSV, visualisations
🔄 Data augmentation (optional)	augment_images.py	On-the-fly augmentations for stage 1, Apply random affine transformation (small rotation + scaling) to a single image.

Methods and Results

Classification Report: Accuracy : 97.50% Precision : 0.9902 Recall : 0.7250 F1 Score : 0.8371 TP=203 | TN=2878 | FP=2 | FN=77 Inference Total pairs: 3160

---

Step-by-step

1. Environment

# Recommended: Python ≥3.9 + venv / conda
pip install -r requirements.txt

or

# Recommended: Python ≥3.9 + venv / conda
conda env create -f environment.yml  
conda activate fvc-fingerprint
cd project3_fingerprint_fvc2000

2. Folder layout (after you run everything once)

project-root/
│
├─ data/
│   ├─ original/                # raw + enhanced images
│   │   ├─ DB1_B/               # raw .tif from the FVC-2000 DB1_B set
│   │   └─ DB1_B_new/           # enhanced .tif created by preprocess.py
│   ├─ processed/               # network-ready .npy files
│   │   └─ DB1_B_new_1/
│   └─ pairs files (*.npz)      # training / val pairs 
│
├─ checkpoints/                 # *.pth saved by train.py
├─ outputs/                     # simple visulization for training and validation loss / metrics
└─ *.py                         # source code

3. Pre-process the raw fingerprints

python preprocess.py \
  --input   ./data/original/DB1_B \
  --out-npy ./data/processed/DB1_B_new_1 \
  --out-tif ./data/original/DB1_B_new_1 \
  --size    300

Advanced pipeline: Gamma → zero-mean/unit-var normalisation → orientation field → overlapping-block Gabor → CLAHE → resize.

The resulting .npy files are float32, shape (300, 300), range 0–1.

4. Build training / validation pairs

python create_train_pairs.py

5. Train

python train.py \
  --train_pairs ./data/*.npz \
  --val_pairs ./data/*.npz \
  --use_aug \
  --balance_neg \
  --finetune \
  --batch_size \
  --lr 1e-3
  

if finetune on certain checkpoints:

python train.py \
  --train_pairs ./data/*.npz \
  --val_pairs ./data/*.npz \
  --use_aug \
  --balance_neg \
  --finetune \
  --best_ckpt ./checkpoints/*.pt 
  

6. Validate

python validate.py \
  --val_data ./data/*.npz \ 
  --ckpt ./checkpoints/*.pt \
  --threshold 1.01

7. Run batch inference on unseen data

With 12% Sample calibration

python inference_batch.py \
  --inference_data ./data/original/DB1_B \
  --ckpt ./checkpoints/model.pt \
  --auto_threshold

---

• Changing the CNN architecture – edit siamese_model.py; keep the output embedding dimension consistent across training and inference.
• Different FVC subsets – simply pass a different --input folder to preprocess.py; pair-generation is fully data-driven.
• GPU vs. CPU – the heavy lifting is in PyTorch; OpenCV Gabor kernels run on CPU. If preprocessing becomes a bottleneck, experiment with smaller block_size or multiprocessing.
• Reproducibility – train.py seeds torch, numpy, and Python’s built-in RNG; pass --deterministic for deterministic CuDNN.