Distance transform data augmentation and stochastic patch-wise image prediction methodology for small dataset learning

Adam Hammoumi (1) , Maxime Moreaud (1, 2) , Christophe Ducottet (3) , Sylvain Desroziers (1)

1 IFPEN - IFP Energies nouvelles
2 CMM - Centre de Morphologie Mathématique
3 LHC - Laboratoire Hubert Curien

Most recent methods of image augmentation and prediction are building upon the deep learning paradigm. A careful preparation of the image dataset and the choice of a suitable network architecture are crucial steps to assess the desired image features and, thence, achieve accurate predictions. We first propose to help the learning process by adding structural information with specific distance transform to the input image data. To handle cases with limited number of training samples (as 12 training and 2 validation images), we propose a patch-based procedure with a stratified sampling method. We illustrate our approaches on image dataset generated by an FFT-based ho-mogeneization technique for heterogeneous media physical properties. The obtained results are evaluated using SSIM, UIQ and PSNR metrics. The proposed techniques demonstrate that the established framework is a reliable estimation method that could be used for a wide range of applications.

Mots clés: image augmentation; deep learning; distance transform; patch-wise segmentation; stratified sampling

Domaines: Intelligence artificielle [cs.AI]; Traitement des images [eess.IV]

Adam Hammoumi, Maxime Moreaud, Christophe Ducottet, Sylvain Desroziers. Distance transform data augmentation and stochastic patch-wise image prediction methodology for small dataset learning. Neurocomputing, inPress. ⟨hal-02879709v1⟩