Skip to Main content Skip to Navigation

Kernel based learning on hierarchical image representations : applications to remote sensing data classification

Yanwei Cui 1, 2
2 OBELIX - Environment observation with complex imagery
Abstract : Hierarchical image representations have been widely used in the image classification context. Such representations are capable of modeling the content of an image through a tree structure, where objects-of-interest (represented by the nodes of the tree) can be revealed at various scales, and where the topological relationship between objects (e.g. A is part of B, or B consists of A) can be easily captured thanks to the edges of the tree. However, for fully benefiting from this key information, dedicated machine learning methods that can directly learn on hierarchical representations and handle the induced structured data need to be developed. In this thesis, we investigate kernel-based strategies that make possible taking input data in tree-structured and capturing the topological patterns inside each structure through designing structured kernels. We apply the designed kernel to remote sensing image classification tasks, allowing the discovery of complex cross-scale patterns in hierarchical image representations. We develop a structured kernel dedicated to unordered tree and path (sequence of nodes) structures equipped with numerical features, called Bag of Subpaths Kernel (BoSK). BoSK is an instance of a convolution kernel relying on subpath substructures, more precisely a bag of all paths and single nodes. It is formed by summing up kernels computed on all pairs of subpaths of the same length between two bags. The direct computation of BoSK can be done through an iterative scheme, yielding a quadratic complexity w.r.t. both structure size (number of nodes) and amount of data (training size). However, such complexity prevents BoSK to be used on real world large-scale problems, where the tree can have more than hundreds of nodes and the available training data can consist in more than ten thousands samples. Therefore, we propose a fast version of the algorithm, called Scalable BoSK (SBoSK for short), using Random Fourier Features to map the structured data in a randomized finite-dimensional Euclidean space, where inner product of the transformed feature vector approximates BoSK. It brings down the complexity from quadratic to linear w.r.t. structure size and amount of data, making the kernel compliant with the large-scale machine learning context. Thanks to (S)BoSK, we can learn from cross-scale patterns in hierarchical image representations. (S)BoSK operates on paths, thus allowing modeling the context of a pixel (leaf of the hierarchical representation) through its ancestor regions at multiple scales. Such a model is used within pixel-based image classification. (S)BoSK also deals with trees, making the kernel able to capture the composition of an object (top of the hierarchical representation) and the topological relationships among its subparts. This strategy allows tile/sub-image classification. Further relying on (S)BoSK, we introduce a novel multi-source classification approach that performs classification directly from a hierarchical image representation built from two images of the same scene taken at different resolutions, possibly with different modalities. Evaluations on several publicly available datasets illustrate the superiority of (S)BoSK compared to state-of-the-art remote sensing classification methods in terms of classification accuracy, and experiments on a urban classification task show the effectiveness of the proposed multi-source classification approach.
Complete list of metadatas

Cited literature [220 references]  Display  Hide  Download
Contributor : Yanwei Cui <>
Submitted on : Friday, December 22, 2017 - 1:07:15 PM
Last modification on : Wednesday, August 5, 2020 - 3:50:41 AM


Files produced by the author(s)


  • HAL Id : tel-01671529, version 1


Yanwei Cui. Kernel based learning on hierarchical image representations : applications to remote sensing data classification. Machine Learning [stat.ML]. Université Bretagne Loire 2017. English. ⟨tel-01671529⟩



Record views


Files downloads