binary images clustering with k-meansdsie10/presentations/session 5/binary images clustering...
TRANSCRIPT
Presentation
João Ferreira Nunes
Instituto Politécnico de Viana do Castelo
Faculdade de Engenharia da Universidade do Porto
DSIE’10 – 2010.01.29
Binary Images Clustering with k-means
Outline
• Introduction
• The Process
– Dataset
– Pre-processing
– Clustering Analysis
• Results
• Conclusions and Future Work
João Ferreira Nunes @ DSIE’10 - FEUP
Introduction
• Develop a method to group binary images inrespect to their content by means of anunsupervised learning technique, k-means;
• Use a set of clustering quality criteria tovalidate the clustering and also to assist theselection of the best number of clusters.
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Image Clustering Process
Data Collection
Pre-processing
Features Extraction
ClusteringImage
ClustersDataset
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
MPEG-7 Core Experiment CE-Shape-1
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Dataset characteristics
• Binary silhouette images that represent objects• Their shape may change due to:
– change of a view point with respect to objects;– non-rigid object motion;– noise resulted from segmentation or digitization;
• Some images have holes, while others do not;• Some images have experienced a number of
transformations, such as scales, distortions, cutsand rotations;
• The size of the images is not constant;
João Ferreira Nunes @ DSIE’10 - FEUP
Introduction Clustering Process Results Conclusions
Image samples
João Ferreira Nunes @ DSIE’10 - FEUP
Introduction Clustering Process Results Conclusions
Experimental Dataset
apple-15 bell-3 heart-10 device3-12 cup-2
spoon-5 bone-13 guitar-7 key-16 hammer-13
horseshoe-5 horse-4 fork-7
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Image Clustering Process
DatasetImage
EnhancementFeatures
ExtractionClustering
Image Clusters
Pre-Processing
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
• Extraction of the ROI (region of interest)
– cropping the images
through their bounding box:
• Noise reduction
– close morphological filter:
Pre-Processing
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Image Clustering Process
DatasetPre-
ProcessingFeatures
ExtractionClustering
Image Clusters
Features Extraction
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Features Extraction
• Solidity:
• Axis Ratio:
• Filled Ratio:
• Perimeter-Area Ratio:
• Eccentricity:
• Extent:
• Invariant moment (skew invariant)
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Area
Convex Hull Area
Minor Axis Length
Major Axis Length
Area
Filled Area
Perimeter
AreaEccentricity of the elipse
Area
Bounding Box Area
Experimental Dataset Features Representation
apple-15 bell-3 heart-10 device3-12 cup-2
spoon-5 bone-13 guitar-7 key-16 hammer-13
horseshoe-5 horse-4 fork-7
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Image Clustering Process
DatasetPre-
ProcessingFeatures
ExtractionClustering Analysis
Image Clusters
Clustering
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
K-means Clustering
• Images are allocated into K different sets,according to their level of similarity (Euclideandistance);
• Minimizes the intra-cluster distance andmaximizes the inter-cluster distance;
• The value of K “should” be known in advance;
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
K-means Clustering – Getting the best K
• Several clustering iterations were conducted,varying k from 3 to 20;
• Computed some internal criteria (with noinformation a priori) that validate eachclustering solution:– Silhouette index
– Calinski-Harabasz index
– C index
– weighted inter-intra index
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Internal criteria
João Ferreira Nunes @ DSIE’10 - FEUP
Introduction Clustering Process Results Conclusions
Image Clustering Process
Data Collection
Pre-Processing
Features Extraction
ClusteringImage
ClustersImage
Clusters
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Results
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
C1 C2 C3 C4 C5 C6apple 20 20bell 20 20
bone 18 2 20cup 12 8 20
device3 15 5 20fork 15 5 20
guitar 1 19 20hammer 8 12 20
heart 19 20horse 20 1 20
horseshoe 20 20key 19 1 20
spoon 8 12 2028 48 54 57 25 48
Results
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Results
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Conclusions and Future Work
• Achieved results are encouraging and suggest the adequacy of the selected features;
• Future Work Explore new features
Consider weighting features
Increase the Dataset
Compare with other clustering methods
Develop a CBIR system using a supervised method (e.g.: k-nearest neighbor)
Introduction Clustering Process Results Conclusions
João Ferreira Nunes @ DSIE’10 - FEUP
Presentation
João Ferreira [email protected]
Thank You! – Questions?
DSIE’10 – 2010.01.29
Binary Images Clustering with k-means