Methodology
Figure 2: The main steps of the proposed method.
The block size as well as the minimum (MINR)
and the maximum (MAXR) number of possible regions are
the three parameters of the proposed method method. The proposed scheme [1] consists of
several steps.
-
The image is divided into overlapping blocks, whose content is described by the distribution of three color Lab components and two components from wavelet transform to represent the texture [8].
-
The Voting framework [4-5] selects the MAXR representative blocks using as input the blocks grid graph or its minimum spanning tree and the Mallows distance (see Fig. 1).
-
Then, for each possible number of regions K ∈ {MINR, ...,MAXR}, we have performed the following steps (see lines 4-9 of Fig. 2). Based on the Mallows distance and the K first representative blocks as extracted by a Voting Framework, the rest blocks are grouped in clusters with homogeneous content. Then a k-centroids clustering procedure is applied using the Bhattacharyya distance between the distributions of features extracted from each block cluster. Based on the resulting clustering, the feature classes are extracted and described, one per cluster, and then, probabilistic distances are used to select and label a set of pixels that belong to each class with high confidence, thus providing an initial map of almost correctly labelled pixels. Different existing components of the same class are captured by the initial seed-regions. Having the data modelling available and the initial map of almost correctly labelled pixels, we have used an algorithm in order to fill the initial map of decisions, using statistical dissimilarity criteria. The new algorithm, Priority Multi-Class Flooding Algorithm (PMCFA) [2,3] imposes strong topology constraints in such a way that also allows topological flexibility. The segmentation map is constructed respecting the proximity principle. The final step of our framework is a statistical region merging algorithm [2].
-
Finally, we select the segmentation that minimizes a criterion taking into account the average likelihood per pixel of the classification map and penalizes the complexity of the regions boundaries.