Publications/crozet.14.icip.inc

Figures

Fig. 1

(a) Denoising (self-dual grain removal).
(b) Shape Filtering (keep round objects).
(c) Object Detection (energy-based method).
	x
(d) Hierarchical Segmentation (saliency-based).
(d') Hierarchical Segmentation: fine (left), coarse (right).

Fig. 1: Sample uses of the tree of shapes.

Fig. 2

(a) Image	(b) Tree of shapes

Fig. 2: An image (a) and its tree of shapes (b). The propagation of the level line λ ended, meaning that the nodes O and A have already been visited. The hierarchical queue contains the interior contour of B and C. Thus it can be partitioned in two sets S⁺λ</math> = ∂B and S⁻λ = ∂C. The propagation can proceed on both parts in parallel.

Fig. 4

	~ ~	~ ~
(a) Input	(b) Subdivided	(c) Immersed

Fig. 4: (a) is the input image. (b) is the result of the subdivision. (c) is the result of the immersion into the Khalimsky grid. 0-faces are represented by dots, 1-faces by segments and 2-faces by squares.

Fig. 8

(a) Original image	(b) Re-valued image

Fig. 8: The original image (a) and the associated F^{ord} (b); the max-tree of (b) coincides with the tree of shapes of (a).

Fig. 10

Fig. 10: Computation times (in seconds) on a classical image test set of the following algorithms: FLLT [3], FLST [23], Géraud et al. [2], and this paper proposal.

Images

Images used for the benchmarks: [1]

Source Code

• Code of the serial version: [2]
• Code of the parallel version: [3]
• Code of the milena image processing library: [4]