DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

23
Detection of Compound Structures Using Hierarchical Clustering of Statistical and Structural Features H. G ¨ okhan Akc ¸ ay Selim Aksoy Department of Computer Engineering Bilkent University Bilkent, 06800, Ankara, Turkey {akcay,saksoy}@cs.bilkent.edu.tr IGARSS 2011 IGARSS 2011 c 2011, Akcay and Aksoy (Bilkent University) 1 / 23

Transcript of DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Page 1: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Detection of Compound Structures UsingHierarchical Clustering of

Statistical and Structural Features

H. Gokhan Akcay Selim Aksoy

Department of Computer EngineeringBilkent University

Bilkent, 06800, Ankara, Turkey{akcay,saksoy}@cs.bilkent.edu.tr

IGARSS 2011

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 1 / 23

Page 2: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Introduction

I A common approach for object recognition is to segmentthe images into homogeneous regions.

I An alternative for semantic image understanding is toidentify the image regions that are intrinsicallyheterogeneous.

I Examples of such regions, also called compoundstructures, include different types of residential,commercial, industrial, and agricultural areas.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 2 / 23

Page 3: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Introduction

I Compound structures are comprised of spatialarrangements of primitive objects such as buildings, roads,and trees.

I Our aim is to model compound structures in a hierarchicalsegmentation framework that can be used for semanticclassification, annotation, indexing, and retrievalapplications.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 3 / 23

Page 4: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Outline

1 Detection of primitive objects

2 Feature extractionImage representationStatistical featuresStructural features

3 Grouping of primitive objects

4 Experiments

5 Conclusions

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 4 / 23

Page 5: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Detection of primitive objects

I The set of primitives includes objects that can be relativelyeasily extracted using low-level operations that exploitspectral, textural, and morphological information.

I The buildings that are used as primitive objects in theexperiments are detected using thresholding of differentspectral bands.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 5 / 23

Page 6: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Detection of primitive objects

(a) Ankara image (b) Detected buildings

Figure 1: Examples of building detection in a WorldView-2 image of Ankara,Turkey. The detected buildings are highlighted with red in (b).

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 6 / 23

Page 7: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionImage representation

I The image content is modeled using attributed relationalgraphs where the primitives form the vertices.

I We assume that neighboring objects can be related, andconnect every neighboring vertex pair with an edge.

I We use a threshold on the distance between the centroidsof object pairs to determine the neighbors.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 7 / 23

Page 8: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionImage representation

(a) Building mask (b) Neighborhood graph

Figure 2: Examples of graph construction. The vertices considered asneighbors based on proximity analysis are connected with red edges in (b).

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 8 / 23

Page 9: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStatistical features

I The statistical features that summarize the spectral contentand the shape of each individual object consist of

I mean values of the pixels within the object for each spectralband,

I area,I eccentricity, andI centroid location.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 9 / 23

Page 10: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

I The structural features represent the spatial layout of eachobject with respect to its neighbors.

I An important structural information is the amount ofalignment among objects.

I We propose a new method for the detection of alignedobject groups.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 10 / 23

Page 11: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

I A group of three or more objects are accepted as aligned iftheir centroids lie on a single straight line.−→ Least-squares line fitting on the centroid locations.

I Another important factor in an alignment is the uniformity ofthe spacing among the objects in the group.−→ Standard deviation of the distances between the centroids.

Figure 3: Illustration of object alignment.IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 11 / 23

Page 12: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

I Given an image containing n objects, groups of alignedones can be found by checking all possible subsets havingat least three objects.

I These subsets can be generated using a depth-first searchon the neighborhood graph.

I Depth-first search starts at a particular vertex v, andrecursively traverses all vertices connected to v.

I This procedure is repeated by starting the search algorithmfrom each vertex in the graph.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 12 / 23

Page 13: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

I Once a path with at least three vertices is obtained, it isconsidered for possible alignment.

I We perform pruning by using thresholds on the sum ofsquared residuals for line fitting and the standard deviationof centroid distances.

I At the end of the search procedure, the set of structuralfeatures computed for each object group consists of

I orientation of the fitted line, andI mean of the centroid distances.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 13 / 23

Page 14: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

(a) Building mask (b) Alignment detection

Figure 4: Examples of alignment detection. All groups of buildings satisfyingthe alignment criteria are shown in (b).

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 14 / 23

Page 15: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

(a) Building mask (b) Alignment detection

Figure 5: Examples of alignment detection. A total of 622 aligned groups ofthree or more objects are detected but only 34 are shown for clarity.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 15 / 23

Page 16: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Feature extractionStructural features

(a) Building mask (b) Alignment detection

Figure 6: Examples of alignment detection. All groups of buildings satisfyingthe alignment criteria are shown in (b).

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 16 / 23

Page 17: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Grouping of primitive objects

I After each vertex is assigned statistical and structuralfeatures, the next step is to group these objects using aniterative multi-level hierarchical clustering.

I The pairwise object distances in hierarchical clustering canbe computed separately for statistical features andstructural features.

I The distance for two objects with respect to statisticalfeatures is computed using sum of squared differencesbetween the corresponding features of these objects.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 17 / 23

Page 18: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Grouping of primitive objects

I The distance with respect to structural features is computedfrom the alignment groups that these objects belong to.

I The distance between two alignment groups is computedas the sum of squared differences between thecorresponding features of these groups.

I The distance for two objects is computed as the minimum ofthe distances between all pairs of alignment groups whereone group in a pair is associated with one of the objects andthe other group is associated with the other object.

I If at least one of the objects is not found to belong to anyalignment group, the distance of that object to any otherobject is set to∞.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 18 / 23

Page 19: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Grouping of primitive objects

I Once the statistical and structural distances are computedfor each neighboring object pair, agglomerative hierarchicalclustering iteratively groups these objects.

I We combine the results of separate clusterings usingstatistical distances and structural distances.

I The two clustering results are combined by assigning theobjects the same label if they are determined to belong tothe same cluster by either statistical distances or structuraldistances.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 19 / 23

Page 20: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Experiments

I We performed experiments on the WV-2 image of Ankara.

I A graph was constructed with 418 vertices corresponding tothe detected buildings and 2, 610 edges obtained usingproximity analysis.

I The alignment detection algorithm resulted in 622 groups ofthree or more buildings aligned within the allowed limits.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 20 / 23

Page 21: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Experiments

I Clustering using statistical distances resulted in 37 groupsof a total of 254 buildings.

I Clustering using structural distances resulted in 80 alignedgroups of a total of 402 buildings.

I The combined clustering produced 38 groups containing atotal of 403 buildings.

I We can conclude that groups of buildings with differentcharacteristics and spatial layouts that cannot be obtainedby traditional segmentation methods are successfullyextracted by the proposed method.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 21 / 23

Page 22: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Experiments

(a) Statistical clustering (b) Structural clustering (c) Combined clustering

Figure 7: Example clustering results. Different groups are shown in differentcolors. The buildings that did not merge to any group at the given level areshown as white.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 22 / 23

Page 23: DetectionOfCompoundStructuresUsingHierarchicalClusteringOfStatisticalAndStructuralFeatures.pdf

Conclusions

I We described a new procedure that combinesI statistical characteristics of primitive objects modeled using

spectral, shape, and position information withI structural characteristics encoded using their spatial

alignments modeled in terms of joint orientation anduniformity of inter-object spacing

for the detection of compound structures.

I Future work includes methods for automatic selection ofparameters and experiments on larger data sets.

IGARSS 2011 c©2011, Akcay and Aksoy (Bilkent University) 23 / 23