PerRFI: Multi-Domain RFI Mitigation & Image Inversion

• PerRFI is a polysemous term defining systems for radio frequency interference mitigation and forensic classification, spanning real-time FPGA excision, PCA/Fourier analysis, and coordinated site monitoring.
• In radio astronomy, implementations like the uGMRT and FAST pipelines employ robust statistical estimators and spectral decomposition to achieve up to 8.7 dB interference rejection and a 5.5× reduction in cross-correlation error bars.
• In computer vision, PerRFI refers to a Piecewise Rectified Flow Inversion method that improves few-step ODE-based latent inversion, yielding higher PSNR and SSIM compared to traditional DDIM inversion.

PerRFI is a polysemous research term rather than a single standardized method. In radio astronomy, it has been used for at least three distinct efforts: a real-time broadband impulsive-RFI excision system in the uGMRT wideband backend, a PCA- and Fourier-based forensic pipeline for categorizing and cataloguing interference in FAST pulsar-search data, and an INAF-led coordinated program for monitoring and mitigating interference at the Medicina, Noto, and Sardinia radio telescopes. In an unrelated computer-vision context, PerRFI denotes “Piecewise Rectified Flow Inversion,” an inversion strategy for few-step text-guided image editing. The term therefore designates context-specific architectures whose commonality is nominal rather than methodological (Buch et al., 2022, Yuan et al., 2022, Zanichelli et al., 2022, Gong et al., 8 Aug 2025).

1. Research usages and disambiguation

The main research usages of PerRFI are summarized below.

Usage of PerRFI	Domain	Core role
Real-time broadband RFI excision/filtering system	uGMRT radio astronomy backend	Detects impulsive broadband RFI and replaces contaminated samples
PCA + Fourier forensic pipeline	FAST pulsar-search data	Categorizes and catalogues impulsive, periodic, and non-stationary RFI
Coordinated monitoring and mitigation project	Italian radio telescopes	Combines site monitoring, hardware tools, and offline flagging
Piecewise Rectified Flow Inversion	Text-guided image editing	Inverts real images in few-step RectifiedFlow editing

In the radio-astronomy literature, the name is consistently attached to interference-centric workflows, but those workflows operate at different stages of the observing chain: real-time excision, offline forensic analysis, and site-level monitoring plus tooling. In the computer-vision literature, by contrast, the same string is an acronym internal to the RectifiedFlow family and has no connection to radio-frequency interference. A common misconception is therefore to treat PerRFI as a single canonical radio-astronomy package; the published record instead indicates a shared label spanning unrelated systems (Buch et al., 2022, Yuan et al., 2022, Zanichelli et al., 2022, Gong et al., 8 Aug 2025).

2. Real-time broadband excision in the uGMRT backend

At the uGMRT, PerRFI is a real-time broadband RFI detection and filtering system implemented in the GMRT Wideband Backend. It targets man-made broadband impulsive interference, especially power-line sparking below $800$ MHz, with the paper identifying Band-2 ($120$–$250$ MHz) as the strongest regime and also noting contamination in Band-3 ($250$–$500$ MHz), Band-4 ($550$–$850$ MHz), and in some cases up to $1$ GHz. The interference is described as typically $10$–$20$ dB above system noise and often periodic at $120$0 ms or $120$1 ms, tracking the $120$2 Hz mains frequency in India (Buch et al., 2022).

The implementation operates on Nyquist-sampled digital time series from each antenna and polarization on an FPGA-based ROACH-1 board. A single board processes $120$3 inputs, each digitized to $120$4-bit precision at $120$5 MHz sampling frequency, and forwards filtered output to the CPU/GPU cluster through $120$6 Gigabit Ethernet. Detection uses a robust statistical estimator, specifically a Median-of-MAD estimator, to set a threshold; contaminated samples are then replaced either by digital noise samples or by a user-defined constant value. Before recording, RF input power is equalized across antennas, and ADC loading is set so that normal noise occupies about up to $120$7 bits of the $120$8-bit converter, leaving headroom for RFI.

A distinctive methodological feature is the simultaneous recording of unfiltered and filtered data through FPGA digital copy modes. In $120$9 digital copy mode, one antenna stream is duplicated, with one copy filtered and the other left unfiltered; in $250$0 mode, one stream is copied to three additional inputs, enabling threshold and replacement-scheme comparisons under identical signal conditions. This comparison framework underlies the commissioning analyses in both beamformer and correlator modes.

The beamformer-mode diagnostics include spectrograms, single-channel time series, and Fourier transforms of channelized data. Broadband power-line RFI appears as periodic broadband bursts lasting about $250$1 ms and recurring every $250$2 ms. In the paper’s example, channel number $250$3, corresponding to $250$4 MHz RF, is used for detailed inspection. The theoretical thermal-noise mean-to-RMS ratio for total intensity is written as $250$5; with $250$6 and $250$7, the expected value is $250$8. Processing gain is defined as

$250$9

The Fourier-domain diagnostic shows suppression of the $250$0 Hz fundamental, the $250$1 Hz harmonic, and higher harmonics, with the paper reporting $250$2 dB rejection at $250$3 Hz.

Correlator-mode validation is based on normalized cross-correlation functions, closure phase, and calibrated visibilities. For a triangle of antennas, the closure relation is

$250$4

in the absence of RFI. In the $250$5 commissioning example of 14 June 2018, using $250$6 antennas, $250$7 MHz processing bandwidth, $250$8 channels, $250$9 s correlator integration, $500$0 beamformer integration, a $500$1 threshold, digital-noise replacement, and an average RFI fraction of $500$2–$500$3 per antenna, the unfiltered closure phase ranged from $500$4 to $500$5, whereas the filtered closure phase stayed within $500$6. The best-case reduction in cross-correlation error bars was a factor of $500$7. The principal caveat is that digital-noise replacement slightly increases scatter on long baselines, quantified as an $500$8 increase in RMS for filtered amplitudes on baselines $500$9 kilowavelengths.

3. Forensic categorization and cataloguing in pulsar-search data

In FAST pulsar-search work, PerRFI is not primarily an excision filter but a forensic classifier intended to categorize, timestamp, and localize interference. Its rationale is that many standard mitigation tools remove contamination without establishing what the interference is, when it appears, or where it originates. The pipeline therefore combines component analysis and temporal analysis to classify three operational categories: impulsive RFI, periodic RFI, and non-stationary RFI (Yuan et al., 2022).

The input is a two-dimensional data block with $550$0 time samples and $550$1 frequency channels; the FAST PSRFITS examples typically use $550$2-bit data with sampling time around $550$3. PerRFI first applies PCA via an SVD-based covariance decomposition. The paper represents the covariance as

$550$4

with the principal spectral bases $550$5 used as frequency-domain components and the associated weights used as time series. Components are retained until the cumulative explained variance exceeds $550$6,

$550$7

Channels whose basis amplitudes exceed $550$8 are treated as narrowband-contaminated.

Temporal discrimination is then applied to the PCA weight curves. An impulsive event is flagged when the weight exceeds $550$9. Periodicity is assessed by removing a quadratic baseline, taking the FFT of the residual, and using a combined threshold built from the mean frequency of peaks above $850$0 and the dispersion of those peaks; periodic RFI is identified when this combined quantity exceeds $850$1 Hz. If a component is neither impulsive nor periodic by these criteria, it is classified as non-stationary. Metadata recorded for each event include time or UTC, frequency channels, telescope pointing in RA/Dec and Alt/Az, and pulse S/N or period.

In simulation, using Gaussian noise backgrounds with $850$2 time samples and $850$3 channels, the classifier identifies narrowband non-periodic RFI above $850$4, narrowband periodic RFI above $850$5, and wideband impulsive RFI above $850$6, with F1 scores between $850$7 and $850$8. The reported real-data classwise metrics are more heterogeneous: periodic RFI has recall $850$9, precision $1$0, and F-score $1$1; narrowband impulsive RFI has recall $1$2, precision $1$3, and F-score $1$4; wideband impulsive RFI has recall $1$5, precision $1$6, and F-score $1$7.

Its main scientific contribution is longitudinal rather than merely local. The pipeline identified a drifting periodic RFI class whose spectral location changes with time, with examples at $1$8 MHz, $1$9 MHz, and $10$0 MHz; such signals are detrimental to pulsar search because they evade static birdy lists. It also found evidence of increased impulsive RFI when FAST points toward the north, in the direction of nearby towns and the city of Guiyang, and reported time-of-day increases associated with morning and midday human activity. PerRFI in this usage is therefore a monitoring and source-tracing framework as much as a classifier.

4. Coordinated site monitoring and mitigation infrastructure in Italy

In the Italian context, PerRFI denotes a coordinated national project involving the Medicina, Noto, and Sardinia Radio Telescope sites. The project combines continuous and campaign-based site monitoring, hardware development for high-resolution RFI detection, and an offline flagging environment. The monitored frequency coverage is $10$1–$10$2 GHz, aligned both with bands allocated to Radio Astronomy Service and Space Research and with the wider ranges used by current and planned receivers (Zanichelli et al., 2022).

The monitoring architecture is site-dependent. Medicina uses a fixed receiving station on a $10$3-m tower with $10$4 azimuth rotation, covering approximately $10$5–$10$6 GHz and $10$7–$10$8 GHz, plus a mobile laboratory covering $10$9–$20$0 GHz. SRT employs the mobile laboratory, a fixed station using the telescope itself as a sensitive receiving antenna, and piggy-back real-time monitoring of the $20$1–$20$2 MHz IF baseband. Noto was first surveyed with the mobile laboratory and later with a portable receiving system covering $20$3 MHz to $20$4 GHz. Dedicated coordinated campaigns were carried out in December 2014 at Noto, in May 2016 at Noto, and in November 2016 at SRT.

The mobile-laboratory protocol is explicitly standardized. The antenna is raised about $20$5 m above ground, the observing range is split into sub-bands, and for each sub-band four azimuth scans are taken: two polarizations under narrow-band and broad-band settings. Measurements use max-hold mode with $20$6 MHz for broad-band or impulsive interference and $20$7 MHz for narrow-band signals. Up to $20$8 spectral points are used, with channel spacing smaller than the RBW. Each panoramic scan lasts about $20$9 s at about $120$00 deg/s, yielding an effective integration time per beam-sized element from roughly $120$01 s down to $120$02 s, depending on antenna beamwidth.

The occupancy results show strong spatial and spectral heterogeneity. At Medicina, reported occupancies include $120$03 in C-low, $120$04 in S, $120$05 in L-high, $120$06 in C-high, $120$07 in L-low, $120$08 in Ku, and $120$09 in X and K. At SRT, the most severe contamination is in P band, with about $120$10 H and $120$11 V occupancy in the full receiver band, followed by L band at about $120$12 H and $120$13 V. At Noto, the 2016 campaign found $120$14 total occupancy in L band, but $120$15 occupancy in the $120$16–$120$17 MHz sub-band owing to radar; later 2021 measurements showed L-band occupancy increasing to $120$18 and S-band occupancy to about $120$19. The project also documented actionable emitter identification, including a broad-band HiperLAN transmitter in the $120$20–$120$21 MHz RAS secondary-use sub-band near Noto, later switched off after triangulation and reporting, and a satellite spurious emission in the $120$22–$120$23 MHz RAS band traced to the LEO satellite OPTOS n.39420.

A major hardware output is a ROACH-based digital spectrometer with two iADC boards, $120$24-bit ADC precision, $120$25 dB dynamic range, $120$26 Gsample/s digitization, $120$27 MHz bandwidth per polarization, $120$28 or $120$29 channels, minimum $120$30 ms dump time, and $120$31 Gigabit Ethernet output. The signal chain uses a polyphase filter bank with $120$32 nodes and $120$33 dB channel isolation, generates $120$34 complex bins, computes total-intensity spectra, and time-stamps data with a $120$35-bit POSIX TIME register. Control is provided by the Python layer WideBand Lowpass GigaBit (WBLGB) over KATCP/TCP-IP. The backend produced a line-to-noise ratio of about $120$36 on the $120$37 GHz W3(OH) maser when calibrated with $120$38, and resolved terrestrial and radar signals at $120$39–$120$40 ms cadence.

The offline software component, Dish Washer, is a modular prototype for single-dish spectropolarimetric RFI inspection and flagging. Implemented as three Python sub-packages plus a C library with OpenMP, it supports manual inspection, flag propagation across feeds and polarizations, storage of flagged regions in FITS tables, and later reconstruction of flagging history. In this usage, PerRFI is best understood as an institutional mitigation ecosystem rather than a single detector.

5. Position within the broader RFI-mitigation taxonomy

Broader radio-science reviews place RFI mitigation into layered taxonomies spanning preventive, pre-detection, pre-correlation, and post-correlation stages, and emphasize the principle of removing interference as early as possible and at the highest possible time and spectral resolution (Baan, 2023, 1711.01978). This broader framework clarifies the relationships among the radio-astronomy usages of PerRFI.

The Italian PerRFI project occupies the preventive and monitoring end of the chain, combining site surveillance, transmitter identification, and hardware preparation for high-resolution observations. The uGMRT implementation operates at real time on digitized time-series streams, closer to the pre-detection or backend-excision layer. The FAST PCA/Fourier system is fundamentally post-detection and forensic, classifying residual contamination in already formed dynamic spectra. This suggests that “PerRFI” in radio astronomy is best interpreted not as one algorithmic lineage but as a label attached to multiple intervention points within the same systems problem.

This multi-stage interpretation is consistent with adjacent work on real-time and streaming constraints. RFIm for the Apertif Radio Transient System is a GPU/OpenCL library designed for $120$41 Gb/s processing, with Time-Domain Sigma Cut and Frequency-Domain Sigma Cut kernels that reduce trigger load while contributing only $120$42 to $120$43 of total AMBER execution time (Sclocco et al., 2020). A separate FPGA implementation on UniBoard uses robust recursive power estimation and Bernoulli detection at an input data rate around $120$44 Gbps for pre-correlation flagging (Cédric et al., 2017). These related systems reinforce the same architectural pressure that the PerRFI literature repeatedly exposes: mitigation must be matched to the observing stage, data rate, and interference morphology.

6. Distinct computer-vision usage: Piecewise Rectified Flow Inversion

Outside radio astronomy, PerRFI denotes “Piecewise Rectified Flow Inversion,” the inversion component of InstantEdit, a few-step text-guided image editing method built on RectifiedFlow. Here the problem is not radio interference but the difficulty of inverting a real image into a latent trajectory that remains faithful when only $120$45–$120$46 function evaluations are available. Standard DDIM inversion is described as inaccurate in the few-step regime because its local-linearity assumption breaks down; PerRFI instead inverts along a piecewise straightened RectifiedFlow trajectory, yielding smaller approximation error (Gong et al., 8 Aug 2025).

The continuous-time model is written as

$120$47

with Euler-style denoising updates and a corresponding inversion rule of the form

$120$48

The method stores intermediate inverted latents and reuses them during regeneration through Inversion Latent Injection. It is further coupled with Disentangled Prompt Guidance and a Canny-conditioned ControlNet to preserve structural cues and suppress artifacts.

The reported ablation compares DDIM inversion with PerRFI. DDIM reconstruction yields Distance $120$49, PSNR $120$50, LPIPS $120$51, MSE $120$52, SSIM $120$53, Whole CLIPScore $120$54, and Edited CLIPScore $120$55. PerRFI improves these to Distance $120$56, PSNR $120$57, LPIPS $120$58, MSE $120$59, SSIM $120$60, Whole CLIPScore $120$61, and Edited CLIPScore $120$62. In this domain, PerRFI is therefore an inversion strategy for few-step ODE-based generation and should not be conflated with the radio-astronomy acronymic family.

The coexistence of these usages makes disambiguation essential. In radio astronomy, PerRFI denotes infrastructures and algorithms for interference monitoring, detection, excision, and classification. In image editing, it denotes a latent inversion scheme on piecewise rectified flow trajectories. The shared abbreviation is coincidental; the underlying objects, assumptions, and performance criteria are entirely different.