Engineering Internship - Deep Learning for Infrared Thermography Antenna Measurements

Anyfields; Toulouse; On-site; EUR 8,900 - 13,000; Full time

Job summary

Anyfields is seeking an intern in Toulouse to develop physics-driven neural networks for restoring infrared thermography maps of antenna fields. You will work across electromagnetics, imaging, and deep learning with the Innovation team, starting as soon as possible for a 6-month period.

Ideal candidates hold a Bac+5 in engineering or applied mathematics, with Python programming and experience in a DL framework.

Qualifications

• Master’s level in engineering, math, ML, or electromagnetics.
• Programming skills (preferably Python) and experience with a DL framework.
• Knowledge of inverse problems or infrared cameras is a plus.

Responsibilities

• Benchmark synthetic and measured datasets, reproduce conventional methods (denoising, drift removal, deconvolution).
• Estimate detector PSF with a neural network constrained by physics of film diffusion and optics.
• Implement self-supervised deep denoising of maps and quantify gains in dB.
• Validate imaging to resolve individual antenna elements in Ku/Ka bands and report findings.

Skills

Deep learning

Signal processing

Team collaboration

Communication skills

Education

Bac+5 - Engineering school / Master in applied mathematics, signal and image processing, machine learning and/or electromagnetics

Tools

Python

PyTorch

Job description

THE POSITION

ANYFIELDS measures antennas by infrared thermography: a thermosensitive film heated by the radiated field is imaged by an infrared camera, giving a 2D map of the electric field, nonintrusive and very close to the antenna. Two limits remain: the maps are noisy (dynamic range of 25 to 30 dB at best, 35 to 40 dB needed to retrieve side lobes and cross-polarization), and they are blurred by the response of the detector (film and optics), whose resolution is not explained by the theory.

This internship develops physics-driven neural networks, trained without any clean reference, to restore the maps, and compares them rigorously with the conventional methods already implemented.

Missions :

• Benchmark synthetic and measured datasets, reproduction of the conventional methods (wavelet denoising, drift removal, deconvolution, slanted edge).

• Estimate PSF of the detector: neural network constrained by the film heat diffusion and the optics, estimated from a few measurements; explain the gap between the theoretical (about 1 mm) and measured (4 to 8 mm) resolution. Compare with conventional methods.

• Implement Denoising by self-supervised deep denoising of the maps, quantified gain in dB towards 35 to 40 dB, ranking of the limiting factors of the setup.

• Validate by test whether the imaging resolves the individual elements of array antennas in the Ku and Ka bands, before and after improvement. Report on the retained methods.

You will work at the crossroads of electromagnetics, imaging and deep learning, with the engineers of the Innovation team.

Location: Toulouse (city centre)

Start: As soon as possible

Duration: 6 months

Profile

• Bac+5 - Engineering school / Master in applied mathematics, signal and image processing, machine learning and/or electromagnetics

• Programming skills (preferably Python), experience with a deep learning framework

• Knowledge of inverse problems or infrared cameras is a plus

• Team spirit, proactive attitude and good communication skills in an interdisciplinary team