Try   HackMD

Sujets 6.1 : Smiths group

Introduction

Presentation psr Clement Fang

  • Ancien Image 2020

Serge Maitrejean

  • Responsable de l'innovation
  • Doctorat en physique

Eric Garrido

  • Doctorat en physique nucleaire

Groupe Smiths

  • Cote en bourse a Londres
  • 4 divisions
  • Smith detection: scanner et detection
    • Detection de choses illicites ou non-declarees

En france:

  • Base a Vitry-sur-Seine
  • A peu pres 200 personnes
  • IA / traitement d'image

Les techs au sein de Smith

Global presence

Vehicle, cargo & mobile screening

High energy X-ray imaging

Which target / threat we are looking for ?

SD Paris Partnerships

Epita est cense etre la

Cigarettes detection

Some big seizures made thanks to our iCmore in the news

iCmore Weapons detection

More in-depth

Truck Radioscopy

  • Imaging with X-Rays but with a scanning principle

  • Pulsed X-ray source: X-Ray pulses (flash) oh three
    \mus
    every 2 or 3 milliseconds
  • One vertical line of detectors/pixels (5 or 20 mm width, 5 mm height): one column of image is recorded for each X-ray pulses
  • Truck speed is limited: < displacement of detector width between 2 pulses (typically 5-7 km/h)
  • Or resolution is bad (large detectors)

A new tech: Matrix detector

  • Multicolumn detector (column)

  • Large resolution improvement (Left one line, Right Matrix detector)

But noting or nobody is perfect

  • First problem: missing part
  • Easy to solve by slowing down the speed but
    • Superposition: le meme point se voit 2 fois

The problem of depth at low speed: rearranging data ?

  • The way of ordering data is depending on the depth where objects are located.. But we don't know the depth ! It's a stereo effect

Ordering data is depending on the depth

  • We have to assume where in depth the object are located, if we are wrong strong artifacts appears

Turning a drawback onto an advantage

  • Minimizing the artifacts
    Finding the depth of the objects and providing a optimum high resolution image

Curent status

  • Proof of concept has been done using energy minimization technics
  • Work on this approach is pursuing
  • A comprehensive set of data has been acquired from which the "exact images" can be extracted
  • We want to test another approach, neural networks and deep learning are good candidates

The work

  • Getting familiarized with the problem (not so easy)
  • Getting familiarized with the current method
  • Initating Matrix Detector Deep Learning process for:
    • Building the best radioscopic planar images
    • Finding the depth where objects are located