MFN Decomposition & Metrics for HRRP

This repository provides a deep-learning-friendly implementation of the MFN decomposition (Mask–Features–Noise) for High-Resolution Range Profiles (HRRP), together with decomposition-aware similarity metrics designed for evaluating generative models (and more generally comparing HRRP signals in a physically meaningful way).

It accompanies the paper:

MFN Decomposition and Related Metrics for High-Resolution Range Profiles Generative Models
(Accepted at RadarConf25)

---

Why MFN?

HRRPs are 1D radar signatures that often contain:

• a target support region (where the target “lives” along the range axis),
• structured low-frequency components (stable, geometry-driven),
• high-frequency residuals / clutter (unstable peaks, noise-like content).

Standard metrics like global MSE or cosine similarity can be dominated by:

• background / empty cells,
• small alignment differences,
• very localized peaks,
• target size effects (different support lengths).

MFN decomposes a range profile into interpretable components and enables metrics that focus on what matters for HRRP comparison.

---

What’s inside

• utils.py
  Core MFN utilities (decomposition + metric helpers).

• demo_decomposition.ipynb
  A ready-to-run notebook demonstrating MFN decomposition and metric computation.

• data/ship_hrrp.pt
  A small demo tensor with 128 HRRP examples of length 512 (range cells).

• assets/
  Two figures already used by the paper / repo:

  • assets/examples_decomp.png
  • assets/plot_best_wwo_mfn.png

---

Visual overview

MFN decomposition examples

Metric behavior (with / without MFN)

The following figure shows both metrics between two data of the same ship blue and two data from different ships red.

---

MFN decomposition

Given a range profile ( RP \in \mathbb{R}^{S} ), MFN builds:

• m: a soft mask over the cells of interest (COI) / target support
• f: a filtered feature component (Gaussian-smoothed / low-frequency) inside the support
• n: a residual component capturing remaining peaks and noise
  [ n = RP - f ]

The goal is not to “denoise everything”, but to separate robust structure from non-robust peaks so that evaluation metrics become more consistent and interpretable.

---

Decomposition-aware metrics

This repo supports comparing signals using MFN outputs, e.g.:

• MSE on features (feature-only MSE, optionally normalized by an estimate of target support length)
• Cosine similarity on features restricted to the (union of) target supports / COIs

These metrics are designed to be:

• less sensitive to empty/background cells,
• less dominated by isolated peaks,
• more meaningful when comparing same-target vs different-target HRRPs.

---

Citation

If you use this code, please cite:

@inproceedings{brient_mfn_hrrp,
  title     = {MFN Decomposition and Related Metrics for High-Resolution Range Profiles Generative Models},
  author    = {Brient, Edwyn and Velasco-Forero, Santiago and Kassab, Rami},
  booktitle = {IEEE Radar Conference (RadarConf)},
  year      = {2025}
}

Acknowledgments

This work used HPC resources from GENCI–IDRIS (Grant 2024-AD011014422R1).