Logarithmic Morphological Neural Nets robust to lighting variations

Guillaume Noyel (1) , Emile Barbier--Renard (1) , Michel Jourlin (1) , Thierry Fournel (1)

1 LHC - Laboratoire Hubert Curien

Morphological neural networks allow to learn the weights of a structuring function knowing the desired output image. However, those networks are not intrinsically robust to lighting variations in images with an optical cause, such as a change of light intensity. In this paper, we introduce a morphological neural network which possesses such a robustness to lighting variations. It is based on the recent framework of Logarithmic Mathematical Morphology (LMM), i.e. Mathematical Morphology defined with the Logarithmic Image Processing (LIP) model. This model has a LIP additive law which simulates in images a variation of the light intensity. We especially learn the structuring function of a LMM operator robust to those variations, namely : the map of LIP-additive Asplund distances. Results in images show that our neural network verifies the required property.

Mots clés: Morphological neural nets; Logarithmic Image Processing; Logarithmic Mathematical Morphology; Robustness to lighting variations; Functional Asplund metric; Mathematical morpholgy; Machine learning / deep learning

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]; Analyse numérique [math.NA]; Traitement du signal et de l'image [eess.SP]

Guillaume Noyel, Emile Barbier--Renard, Michel Jourlin, Thierry Fournel. Logarithmic Morphological Neural Nets robust to lighting variations. DGMM 2022, IAPR Second International Conference on Discrete Geometry and Mathematical Morphology, Oct 2022, Strasbourg, France. pp.462-474, ⟨10.1007/978-3-031-19897-7_36⟩. ⟨hal-03645285v2⟩