Thèse : Self-Organizing Map Quantization Error Approach for Detecting Temporal Variations in Image Sets

Soutenance de thèse : John WANDETO

Titre : Self-Organizing Map Quantization Error Approach for Detecting Temporal Variations in Image Sets

Résumé : A new approach for image processing, dubbed SOM-QE, which exploits the quantization error (QE) from self-organizing maps (SOM) was put to the test in this work. SOM produce low-dimensional discrete representations of high-dimensional image input data. The QE is determined by the image input data and the resulting final synaptic weights from the SOM's unsupervised learning process. The QE from analyses of time-series of images generates an indicator of potentially critical data change. For implementing SOM, a map size, neuronal neighbourhood distance, learning rate, and number of iterations in the learning process need to be determined. The combination of these parameters returning the lowest QE is always taken as the optimal parameter set for further analysis. The novelty of this approach is fourfold: First, it employs a SOM to determine the QE for different images of a single scene - typically from a time series dataset - unlike the traditional approach where different SOMs are applied on one and the same image. Second, the QE value is then a valid measure of spatial variability across images (proof of concept is provided in the thesis on the basis of several experiments including a detection experiment with human expert and non-expert observers using the concepts of signal detection theory). Third, the QE is shown to generate a specific and unique label for a given image within the dataset. Fourth, this label can be effectively used to track changes that occur in subsequently taken images of one and the same scene, providing a statistically reliable measure of variation in the image data as a function of time. The approach was applied here to computer generated images, medical images, and satellite images. Furthermore, the QE is shown to be reliably correlated with relevant geographic and demographic data. The technique introduced here is fast in terms of overall computation time (in seconds) and economic in terms of implementation. It provides a truly novel and, in principle, simple solution for automatically detecting the evolution of a pathology, or changes in urban and natural landscapes in image data. These may not be visible to the eye of the human expert. Moreover, Pearson's correlation test confirms a statistically significant correlation between the QE metric and image ground truth data.

Le jury est composé de Birgitta DRESP, Directeur de Recherche CNRS (Directeur de thèse pour la France), Henry NYONGESA, Professor of Computer Science and Systems Engineering DEKUT (Directeur de thèse pour le Kenya), Rachele ALLENA, HDR ENSAM (rapporteur), Paul ROSIN, Professor of Computational Science U Cardiff (rapporteur), José RAGOT, Professeur Eméritus en Automatique U Nancy (examinateur), Daniel GEORGE, Maître de Conférences HDR (examinateur) - Membres invités: Yves REMOND, Professeur des Universités.

La soutenance se tiendra en anglais le Vendredi 14 Septembre 2018 2018 à 14h15 en Salle des Conseils au 2, rue Boussingault à Strasbourg.

Elle a été réalisée dans le cadre d'une cotutelle Université de Strasbourg, France - Dedan Kimathi University of Technology, Kenya avec le soutien de l'Ambassade de France à Nairobi et Campus France.