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: SegmentationTime (seconds) Region Growing Watershed Transformation Region Merging

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

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 a small Grid, we never had scalability issues Most Grids are very large and having a single root node is a bottleneck The next stage is to test out a hierarchical model

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

Questions ?