bil717 evrenkaya sertackaya project presentation 2012
TRANSCRIPT
BIL717 Image Processing Course ProjectSerta Bura Kaya Zahit Evren Kaya
Comparison of Image Segmentation Algorithms: Segmentation by Weighted Aggregation (SWA) and gpb-owt-ucm
Image Segmentation Problem Summary
Process of clustering pixels into salient image regions Regions corresponding to individual objects, surfaces, natural parts Label each pixel Pixels with same label share common characteristics
Image Segmentation Problem Summary
Simplify/change the representation of an image Easier to analyze First step of more complex image processing tasks Obtain results close to human visual system
Image Segmentation Common Approaches
PDE-based methods
Snakes, Level Set Normalized Cut
Graph-based methods
Contour-based methods
gPb-ucm-owt, cannyucm-owt
Thresholding Region based methods
Splitting, merging
Compared Methods gPb-owt-ucm
Global Probability of Boundary (gPb): Determine how likely that one point belongs to a boundary Oriented Watershed Transform (owt): Convert probability scores to line segments Ultrametric Contour Map: Hierarchical segmentation from contours
Compared Methods SWA
Segmentation by Weighted Aggregation Graph-based Automated method, no predefinition of size or number of categories Hierarchical approach reduces complexity Utilizes resemblence in luminance
Benchmark Berkeley Segmentation Dataset 500 (BSDS500)
12000 hand labeled segmentations constitues the ground-truth Segmentation quality will be compared Boundary map of a segmentation result is compared to ground-truth Multi-scale thresholding Precision-Recall framework
gPb-owt-ucm Test Results
MATLAB Code and executables are available on Berkeley CV Group website (MATLAB Mex files for Unix platform) Automated and interactive options are available Interactive segmentation run without problems
gPb-owt-ucm Test Results: gPb-owt
gPb-owt-ucm Test Results: gPb-owt
gPb-owt-ucm Test Results: gPb-owt-ucm
gPb-owt-ucm Test Results: gPb-owt-ucm
Problems arise in automated segmentation Eigen-value decomposition is costly Out-of-memory Two methods for approximation:
Stochastic Algorithm Clustered Low Rank Approximation
Graph Based Methods Summary
All pixels of image nodesEdge is formed between nodes. Weight of node Similarity between pixels The image is partitioned into separate sets by removing the edges connecting the segments Partition Method === New Graph Based Segmentation Method
Graph Based Segmentation Summary Graph Based G=(V,E) where V set of n nodes Vi(i=1,2,3,4, ,n) and E set of undirected weighted edges of wij, connecting neighbouring nodes vi, vj Wi,j =e^[- (Ii-Ij)] where Ii and Ij are intensity W matrix is symetric since Wii=0
Pixel Graph
Normalized-Cut Measure
Normalized-Cut Measure
Minimizin Problem creates Sailent Segments where similarity across its boundaries is small in segment is largec
Decreasing Computational Cost of Normalized-Cut Measure
Weighted Aggregation In addition to significantly reducing the number of nodes in the graph, this coarsening creates small aggregates of pixels adapted to the image at hand, the intensities of which are similar. Every pixel belongs to either one or several aggregates, each centered at one seed,
Weighted Aggregation
Hierarchy in SWA