Fast Pattern Spectra using Tree Representation of the Image for Patch Retrieval

Behzad Mirmahboub (1) , Jérôme Moré (1) , David Youssefi (2) , Alain Giros (2) , François Merciol (1) , Sébastien Lefèvre (1)

1 OBELIX - Observation de l’environnement par imagerie complexe
2 CNES - Centre National d'Études Spatiales [Toulouse]

We extend the notion of content based image retrieval to patch retrieval where the goal is to find the similar patches to a query patch in a large image. Naive searching for similar patches by sequentially computing and comparing descriptors of sliding windows takes a lot of time in a large image. We propose a novel method to compute descriptors for all sliding windows independent from number of patches. We rely on tree representation of the image and exploit the histogram nature of pattern spectra to compute all the required descriptors in parallel. Computation time of the proposed method depends only on the number of tree nodes and is free from query selection. Experimental results show the effectiveness of the proposed method to reduce the computation time and its potential for object detection in large images.

Mots clés: content based image retrieval; patch retrieval; tree representation; pattern spectra; large satellite images

Domaines: Traitement des images [eess.IV]

Behzad Mirmahboub, Jérôme Moré, David Youssefi, Alain Giros, François Merciol, et al.. Fast Pattern Spectra using Tree Representation of the Image for Patch Retrieval. DGMM 2021 - IAPR International Conference on Discrete Geometry and Mathematical Morphology, May 2021, Uppsala / Virtual, Sweden. pp.107-119, ⟨10.1007/978-3-030-76657-3_7⟩. ⟨hal-03354933⟩