nabla-dna

Image fingerprinting system combining gradient-based feature extraction (C) with metric learning (Python / Keras).

About

Codebase from an undergraduate ML research internship. The original system was developed by the research lab; this repository preserves the implementation I worked with for portfolio reference. The project explores image fingerprinting via two complementary pipelines:

1. Classical pipeline (C + Python) — Uses nabla (gradient) operators on bitmap images to derive feature histograms.
2. Deep learning pipeline — MobileNet backbone + Dense embedding layer trained with triplet loss to produce a 64-dimensional image fingerprint ("DNA").

Two embeddings can then be compared via cosine similarity for image retrieval and matching tasks.

Tech Stack

Layer	Technology
Native processing	C, CMake
Numerical / image	NumPy, Pillow, SciPy
Deep learning	TensorFlow / Keras, MobileNet
Metric learning	Triplet loss (margin = 0.4, embedding dim = 64)

Project Structure

nabla-dna-master/
├── CMakeLists.txt          # Builds the `mkdna` C executable
├── src/mkdna/              # C source: bitmap → DNA fingerprint
├── python/                 # Python toolchain (BMP, histogram, distance, sobel, ...)
│   ├── mkdna.py            # Python driver
│   ├── nbmp.py, sobel.py, histo.py
│   ├── dist.py, getdist.py, dist_histo.py
│   └── ...
└── DL-DNA/                 # Deep learning pipeline
    ├── dl_dna.py           # CLI entry point
    ├── dl_dna_model.py     # Abstract model base + cosine similarity
    ├── triplet_model.py    # Triplet-loss MobileNet (anchor / positive / negative)
    ├── mobilenet.py        # Plain MobileNet baseline
    └── lineEnumerator.py

Build

Native (C)

cd nabla-dna-master
mkdir build && cd build
cmake ..
make            # produces the `mkdna` executable

Deep learning pipeline

cd nabla-dna-master/DL-DNA
pip install -r requirements.txt

Usage (DL pipeline)

# Train with triplet loss, save the trained model
python dl_dna.py -m triplet_loss -t triples.txt -s model.keras -f images/

# Get the embedding for a single image
python dl_dna.py -l model.keras some_image_name

# Compute cosine similarity between two images
python dl_dna.py -l model.keras image_a image_b

# Compute similarities for a batch (pairs file)
python dl_dna.py -l model.keras pairs.txt

Environment variables: N_UNITS, IMAGE_FOLDER, EPOCHS.

Notes

This repository is a snapshot of the codebase from the internship and is preserved here for portfolio reference. Module naming and architecture reflect the original lab conventions.