Thesis defense of Daniel Maestro Watson

Back

Thesis defense of Daniel Maestro Watson

THESIS

Thesis defense of Daniel Maestro Watson

Title of the thesis:“3D Inspection Methods for Specular or Partially Specular Surfaces”. Obtained the SOBRESALIENTE qualification and he has received the DOCTOR INTERNACIONAL mention

2020·04·23

$titulo.getData()


  • Title of the thesis: “3D Inspection Methods for Specular or Partially Specular Surfaces”.
  • Court:
    • President: Viviane Thérèse Marie Cadenat (LAAS-CNRS)
    • Vocal:  Dimitrios Chrysostomos Chrysostomou (Aalborg University)
    • Vocal: Aritz Legarda Cristobal (Das-nano)
    • Vocal: Bertrand Laurent Aurélien Vandeportaele (UNIVERSITÉ PAUL SABATIER)
    • Secretary: Luka Eciolaza Echeverria (Mondragon Unibertsitatea)

Abstract

Deflectometric techniques are a powerful tool for the automated quality control of specular or shiny surfaces. These techniques are based on using a camera to observe a reference pattern reflected on the surface under inspection, exploiting the dependence of specular reflections on surface normals to recover shape information from the acquired images. Although deflectometry is already used in industrial environments such as the quality control of lenses or car bodies, there are still some open problems. On the one hand, using quantitative deflectometry, the normal vector field and the 3D shape of a surface can be obtained, but these techniques do not yet take full advantage of their local sensitivity because the achieved global accuracies are affected by calibration errors. On the other hand, qualitative deflectometry is used to detect surface imperfections without absolute measurements, exploiting the local sensitivity of deflectometric recordings with reduced calibration requirements. However, this qualitative approach requires further processing that can involve a considerable engineering effort, particularly for aesthetic defects which are inherently subjective.

The first part of this thesis aims to contribute to a better understanding of how deflectometric setups and their calibration errors affect quantitative measurements. Different error sources are considered including the camera calibration uncertainty and several non-ideal characteristics of LCD screens used to generate the light patterns. Experiments performed using real measurements and simulations show that the non-planarity of the LCD screen and the camera calibration are the dominant sources of error. The second part of the thesis investigates the use of deep learning to identify geometrical imperfections and texture defects based on deflectometric data. Two different approaches are explored to extract and combine photometric and geometric information using convolutional neural network architectures: one for automated classification of defective samples, and another one for automated segmentation of defective regions in a sample. The experimental results in a real industrial case study indicate that both architectures are able to learn relevant features from deflectometric data, enabling the classification and segmentation of defects based on a dataset of user-provided examples.