Change Detection and Image Time-Series Analysis 1

4Unsupervised Functional Information Clustering in Extreme Environments from Filter Banks and Relative Entropy

Abdourrahmane M. ATTO¹, Fatima KARBOU², Sophie GIFFARD-ROISIN³ and Lionel BOMBRUN⁴

¹University Savoie Mont Blanc, Annecy, France

²Université Grenoble Alpes, University of Toulouse, CNRM, CNRS, Centre d’Études de la Neige at Météo-France, Grenoble, France

³Université Grenoble Alpes, France

⁴University of Bordeaux, France

This chapter addresses feature extraction from wavelets and convnet (Convolutional Neural Network) filters for unsupervised image time series analysis. We exploit the ability of wavelets and neuro-convolutional filters to capture non-trivial invariance properties, as well as the new centroid solutions proposed in this chapter, for high-level relative entropy-based feature analysis. Anomaly detection and functional evolution clustering are developed from this framework.

4.1. Introduction

Convolutional neural networks (CNNs, convnets) provide new filters for feature extraction in images. In supervised learning approaches, labeled examples of object/texture classes are available in quantity and quality. The main challenges are the design of convnets, the selection of convnet operators and the optimization of convnet parameters. In this context, several scenarios associated with a large class of training databases have made it possible to discover suitable network configurations (Krizhevsky et al. 2012, pp. 1097–1105; Jia et al. 2014; Simonyan and Zisserman ...