arXiv subjects

Francesco Paci

Publications and source records attributed to Francesco Paci.

10 recordsLinked to original sources

arXiv2026

Low-Power License Plate Detection and Recognition on a RISC-V Multi-Core MCU-Based Vision System

In this paper, we present the first (to the best of our knowledge) demonstration of a low-power MCU-based edge device for Automatic License Plate Recognition (ALPR). The design leverages on a 9-core RISC-V processor, GAP8, coupled with a QVGA ultra-low-power greyscale imager. The proposed visual processing pipeline uses a multi-model inference approach based on SSDlite-MobilenetV2 for license plate detection and LPRNet for optical character recognition, reaching a 38.9% mAP score for the first task and a recognition rate of >99.13% for the latter on public datasets. On real-world data, the pipeline recognizes registration numbers when the size of LP crops is as small as 30x5 pixels. Thanks to the applied compression and optimization strategies, the multi-model inference (687 MMAC) achieves a throughput of 1.09 FPS at a power cost of 117 mW when running on GAP8. Our solution is the first MCU-class device embedding such a level of network complexity, resulting to be 73x more energy-efficient w.r.t. precedent mobile-class ALPR system featuring a Raspberry Pi3. The proposed design does not resort to any hardwired acceleration engines, thus retaining full flexibility for future algorithmic improvements.

cs.CV↗

arXiv2024

Self-Learning for Personalized Keyword Spotting on Ultra-Low-Power Audio Sensors

This paper proposes a self-learning method to incrementally train (fine-tune) a personalized Keyword Spotting (KWS) model after the deployment on ultra-low power smart audio sensors. We address the fundamental problem of the absence of labeled training data by assigning pseudo-labels to the new recorded audio frames based on a similarity score with respect to few user recordings. By experimenting with multiple KWS models with a number of parameters up to 0.5M on two public datasets, we show an accuracy improvement of up to +19.2% and +16.0% vs. the initial models pretrained on a large set of generic keywords. The labeling task is demonstrated on a sensor system composed of a low-power microphone and an energy-efficient Microcontroller (MCU). By efficiently exploiting the heterogeneous processing engines of the MCU, the always-on labeling task runs in real-time with an average power cost of up to 8.2 mW. On the same platform, we estimate an energy cost for on-device training 10x lower than the labeling energy if sampling a new utterance every 6.1 s or 18.8 s with a DS-CNN-S or a DS-CNN-M model. Our empirical result paves the way to self-adaptive personalized KWS sensors at the extreme edge.

cs.SD↗

arXiv2022

Accelerating RNN-based Speech Enhancement on a Multi-Core MCU with Mixed FP16-INT8 Post-Training Quantization

This paper presents an optimized methodology to design and deploy Speech Enhancement (SE) algorithms based on Recurrent Neural Networks (RNNs) on a state-of-the-art MicroController Unit (MCU), with 1+8 general-purpose RISC-V cores. To achieve low-latency execution, we propose an optimized software pipeline interleaving parallel computation of LSTM or GRU recurrent blocks, featuring vectorized 8-bit integer (INT8) and 16-bit floating-point (FP16) compute units, with manually-managed memory transfers of model parameters. To ensure minimal accuracy degradation with respect to the full-precision models, we propose a novel FP16-INT8 Mixed-Precision Post-Training Quantization (PTQ) scheme that compresses the recurrent layers to 8-bit while the bit precision of remaining layers is kept to FP16. Experiments are conducted on multiple LSTM and GRU based SE models trained on the Valentini dataset, featuring up to 1.24M parameters. Thanks to the proposed approaches, we speed-up the computation by up to 4x with respect to the lossless FP16 baselines. Differently from a uniform 8-bit quantization that degrades the PESQ score by 0.3 on average, the Mixed-Precision PTQ scheme leads to a low-degradation of only 0.06, while achieving a 1.4-1.7x memory saving. Thanks to this compression, we cut the power cost of the external memory by fitting the large models on the limited on-chip non-volatile memory and we gain a MCU power saving of up to 2.5x by reducing the supply voltage from 0.8V to 0.65V while still matching the real-time constraints. Our design results 10x more energy efficient than state-of-the-art SE solutions deployed on single-core MCUs that make use of smaller models and quantization-aware training.

cs.SD↗

arXiv2015

Supervoids in the WISE-2MASS catalogue imprinting Cold Spots in the Cosmic Microwave Background

The Cold Spot (CS) is a clear feature in the Cosmic Microwave Background (CMB); it could be of primordial origin, or caused by a intervening structure along the line of sight. We identified a large projected underdensity in the recently constructed WISE-2MASS all-sky infrared galaxy catalogue aligned with the Cold Spot direction at $(l,b)\approx(209^\circ,-57^\circ)$. It has an angular size of tens of degrees, and shows a $\sim20\%$ galaxy underdensity in the center. Moreover, we find another large underdensity in the projected WISE-2MASS galaxy map at $(l,b)\approx(101^\circ,46^\circ)$ (hereafter Draco Supervoid), also aligned with a CMB decrement, although less significant than that of the CS direction. Motivated by these findings, we develop spherically symmetric Lemaitre-Tolman-Bondi (LTB) compensated void models to explain the observed CMB decrements with these two underdensities, or "supervoids". Within our perturbative treatment of the LTB voids, we find that the Integrated Sachs-Wolfe and Riess-Sciama effects due to the Draco Supervoid can account for the CMB decrement observed in the same direction. On the contrary, the extremely deep CMB decrement in the CS direction is more difficult to explain by the presence of the CS supervoid only. Nevertheless, the probability of a random alignment between the CS and the corresponding supervoid is disfavored, and thus its contribution as a secondary anisotropy cannot be neglected. We comment on how the approximations used in this paper, in particular the assumption of spherical symmetry, could change quantitatively our conclusions and might provide a better explanation for the CMB CS.

astro-ph.CO↗

arXiv2014

A comparison of CMB Angular Power Spectrum Estimators at Large Scales: the TT case

In the context of cosmic microwave background (CMB) data analysis, we compare the efficiency at large scale of two angular power spectrum algorithms, implementing, respectively, the quadratic maximum likelihood (QML) estimator and the pseudo spectrum (pseudo-Cl) estimator. By exploiting 1000 realistic Monte Carlo (MC) simulations, we find that the QML approach is markedly superior in the range l=[2-100]. At the largest angular scales, e.g. l < 10, the variance of the QML is almost 1/3 (1/2) that of the pseudo-Cl, when we consider the WMAP kq85 (kq85 enlarged by 8 degrees) mask, making the pseudo spectrum estimator a very poor option. Even at multipoles l=[20-60], where pseudo-Cl methods are traditionally used to feed the CMB likelihood algorithms, we find an efficiency loss of about 20%, when we considered the WMAP kq85 mask, and of about 15% for the kq85 mask enlarged by 8 degrees. This should be taken into account when claiming accurate results based on pseudo-Cl methods. Some examples concerning typical large scale estimators are provided.

astro-ph.CO↗

arXiv2013

PRISM (Polarized Radiation Imaging and Spectroscopy Mission): A White Paper on the Ultimate Polarimetric Spectro-Imaging of the Microwave and Far-Infrared Sky

PRISM (Polarized Radiation Imaging and Spectroscopy Mission) was proposed to ESA in response to the Call for White Papers for the definition of the L2 and L3 Missions in the ESA Science Programme. PRISM would have two instruments: (1) an imager with a 3.5m mirror (cooled to 4K for high performance in the far-infrared---that is, in the Wien part of the CMB blackbody spectrum), and (2) an Fourier Transform Spectrometer (FTS) somewhat like the COBE FIRAS instrument but over three orders of magnitude more sensitive. Highlights of the new science (beyond the obvious target of B-modes from gravity waves generated during inflation) made possible by these two instruments working in tandem include: (1) the ultimate galaxy cluster survey gathering 10e6 clusters extending to large redshift and measuring their peculiar velocities and temperatures (through the kSZ effect and relativistic corrections to the classic y-distortion spectrum, respectively) (2) a detailed investigation into the nature of the cosmic infrared background (CIB) consisting of at present unresolved dusty high-z galaxies, where most of the star formation in the universe took place, (3) searching for distortions from the perfect CMB blackbody spectrum, which will probe a large number of otherwise inaccessible effects (e.g., energy release through decaying dark matter, the primordial power spectrum on very small scales where measurements today are impossible due to erasure from Silk damping and contamination from non-linear cascading of power from larger length scales). These are but a few of the highlights of the new science that will be made possible with PRISM.

astro-ph.CO↗

arXiv2013

Constraints on cosmological birefringence energy dependence from CMB polarization data

We study the possibility of constraining the energy dependence of cosmological birefringence by using CMB polarization data. We consider four possible behaviors, characteristic of different theoretical scenarios: energy-independent birefringence motivated by Chern-Simons interactions of the electromagnetic field, linear energy dependence motivated by a 'Weyl' interaction of the electromagnetic field, quadratic energy dependence, motivated by quantum gravity modifications of low-energy electrodynamics, and inverse quadratic dependence, motivated by Faraday rotation generated by primordial magnetic fields. We constrain the parameters associated to each kind of dependence and use our results to give constraints on the models mentioned. We forecast the sensitivity that Planck data will be able to achieve in this respect.

astro-ph.CO↗

arXiv2009

The full contribution of a stochastic background of magnetic fields to CMB anisotropies

We study the contribution of a stochastic background (SB) of primordial magnetic fields (PMF) on the anisotropies in temperature and polarization of the cosmic microwave background radiation (CMB). A SB of PMF modelled as a fully inhomogeneous component induces non-gaussian scalar, vector and tensor metric linear perturbations. We give the exact expressions for the Fourier spectra of the relevant energy-momentum components of such SB, given a power-law dependence parametrized by a spectral index $n_B$ for the magnetic field power spectrum cut at a damping scale $k_D$. For all the values of $n_B$ considered here, the contribution to the CMB temperature pattern by such a SB is dominated by the scalar contribution and then by the vector one at higher multipoles. We also give an analytic estimate of the scalar contribution to the CMB temperature pattern.

astro-ph↗

arXiv2009

Impact of Galactic polarized emission on B-mode detection at low multipoles

We use a model of polarized Galactic emission developed by the the Planck collaboration to assess the impact of foregrounds on B-mode detection at low multipoles. Our main interest is to applications of noisy polarization data and in particular to assessing the feasibility of B-mode detection by Planck. This limits the complexity of foreground subtraction techniques that can be applied to the data. We analyze internal linear combination techniques and show that the offset caused by the dominant E-mode polarization pattern leads to a fundamental limit of r approximately 0.1 for the tensor-scalar ratio even in the absence of instrumental noise. We devise a simple, robust, template fitting technique using multi-frequency polarization maps. We show that template fitting using Planck data alone offers a feasible way of recovering primordial B-modes from dominant foreground contamination, even in the presence of noise on the data and templates. We implement and test a pixel-based scheme for computing the likelihood function of cosmological parameters at low multipoles that incorporates foreground subtraction of noisy data.

astro-ph.CO↗

arXiv2008

The Impact of Stochastic Primordial Magnetic Fields on the Scalar Contribution to Cosmic Microwave Background Anisotropies

We study the impact of a stochastic background of primordial magnetic fields on the scalar contribution of CMB anisotropies and on the matter power spectrum. We give the correct initial conditions for cosmological perturbations and the exact expressions for the energy density and Lorentz force associated to the stochastic background of primordial magnetic fields, given a power-law for their spectra cut at a damping scale. The dependence of the CMB temperature and polarization spectra on the relevant parameters of the primordial magnetic fields is illustrated.

astro-ph↗