Grid-Based Genetic Algorithm Approach to Colour Image Segmentation - PowerPoint PPT Presentation

Grid-Based Genetic Algorithm Approach to Colour Image Segmentation Marco Gallotta Keri Woods Supervised by Audrey Mbogho

Image Segmentation  Identifying and extracting distinct, homogeneous regions from an image  Simplifies the image for further processing: − shape recognition, medical imaging, face detection

Image Segmentation  Problem: How do we segment the following? − Each petal as a region? − Stigma as a region? − Group flowers as single region? − Segment the background?

Human Segmentation  Segmentation by human candidates  Results confirm no single solution

Genetic Algorithms  Optimisation technique that works on large search spaces  Biological evolution

Genetic Algorithms: Chromosome  Chromosome encodes a potential solution  Contains parameters  The chromosome is optimised using: − mutation crossover

Genetic Algorithms: Fitness  Fitness function evaluates an individual and assigns a numerical value  Used to select fittest individuals for next iteration  Crucial in producing good results

Grid Computing  A system that coordinates resources that are not subject to centralized control  Dedicated and non-dedicated resources  Multiple organisations pooling their unused resources  Lots of computing power

Grid Computing

Problem: Segmentation  Segmentation of great importance  No general method of image segmentation  Wide variety of images  Parameters need to be tuned to get optimal results

Problem: Genetic Algorithms  Segmentation involves much uncertainty  GA cope well with uncertainty  Alter parameters to optimise segmentation results

Problem: Computating Requirements  Image segmentation and genetic algorithms computationally intensive  Combined VERY computationally expensive  Solution? − Work harder − Work smarter − Get help

Problem: GA For The Grid  Genetic algorithms easily parallelisable  Grid supplies “free” computational resources

Problem: Research Existing Techniques  Edge detection  Histogram thresholding  Watershed  Region based techniques  Clustering techniques  Model based techniques  Many others

Segmentation Method Implemented  Chose to implement: − Watershed − Region Growing − Region Merging

Watershed Transformation  Calculate a gradient magnitude image  Consider this as a topographic surface  Consider dropping water at each pixel and observing where the trickle ends  Pixels with the same end point form a region

Watershed Transformation  (a) Example gradient magnitude image  (b) The two regions that are identified

Watershed Transformation: Example

Region Growing  Start off with small regions and grow them  Each iteration considers all pixels neighbouring the regions  Pixel with the minimum δ is added to the region  This continues until all pixels are assigned to a region

Region Growing  The above method requires manual seeds  To automate we introduce a threshold T  If the minimum δ exceeds T then a new region is created  Start with an arbitrary pixel as the first region and iterate as above

Region Growing: Example

Region Merging  Initially each pixel a region  Adjacent regions merged if criteria met  Continue until no regions meet criteria

Merging Criterion  Merge if fusion factor less than scale parameter  Fusion factor: change in heterogeneity if regions merged  Heterogeneity: colour, compactness, smoothness  Scale parameter controls size of resulting regions

Region Merging: Example

Segmentation Results • Berkeley Segmentation Dataset • Watershed fastest • Performance results: Segmentation Time (seconds) Region Growing 71.806 Watershed 2.106 Transformation Region Merging 58.059

Segmentation Results

Segmentation Results

Segmentation Results  All successful but different results  Effect of scale parameter on region merging − Large scale parameter => large regions

Segmentation Results: Effect of Scale Parameter

Segmentation Results  [Can get some results off website at http://people.cs.uct.ac.za/~mgallott/honsproj/]

Genetic Algorithm • Modify parameters of region merging algorithm • Scale parameter, weights of components of heterogeneity

GA: Fitness Function  Drives evolution of chromosomes  Evaluate quality of segmentation  Unsupervised segmentation − No external information − Properties of image itself  How much colour within each region varies  Low fitness = good segmentation

GA: Fitness Function  For each region standard deviation multiplied by area  Sum all regions  Add 1  Multiply by number of regions

Genetic Algorithm Results • Inconclusive • Sometimes improvement

Genetic Algorithm Results  GA with segmentation very computationally intensive  Unable to explore full potential  Extremely slow  Therefore grid

Parallel Genetic Algorithms  Two common models: − master-slave (left) − Island model (right)

Grid Computing + Genetic Algorithms  With the Grid, communication between nodes is expensive (“impossible” in a true Grid)  Even with communication, building a topology for the Island model is difficult  All existing research has used the master-slave model

Grid Model  Our model uses ideas from both master-slave and Island models  Root node (dedicated resource) stores a super population  No direct communication between sub-nodes

Grid Model: Results  We were heavily restricted in testing and could only test with eight nodes  Tests showed the communication overhead had negligible impact as fitness function increased in complexity  Results were positive when testing on simple problems  Unsuccessful at migrating the segmentation algorithm to the Grid

Conclusion  Experimented with 3 segmentation algorithms  Selected region merging for our genetic algorithm solution  Genetic algorithm provides potential for improvement but results inconclusive  Grid computing showed positive results however limited resources did not allow for thorough testing

Future Work: Grid  As we only tested on  The next stage is to a small Grid, we test out a hierarchical never had scalability model issues  Most Grids are very large and having a single root node is a bottleneck

Future Work: GA  Watershed with genetic algorithm  Investigate different fitness functions  Genetic programming to evolve fitness function − Train for each desired application

Questions ? http://people.cs.uct.ac.za/~mgallott/honsproj/