|
|
Hollingworth et al. [41] |
Gudmundsson et al. [39] |
Ross et al. [42] | SMART |
|
Application | Generic images (fairly simple) | Unfragmented, localized thin edges in medical images. | Microscopic images from mineral samples. | Generic (satellite images, uniform patterns, etc.). |
Methodology | Exploit inherent parallelism in images | Split image into linked subimages. Maintain links between adjacent pixels. | Implement a training stage (requires sampling 23.6% of image), followed by genetic programming. | Evolve a subset of the edge detector (i.e., critical LUTs) to recover from faults. |
Fitness evaluation | Software model | Software model | Software model | Intrinsic evolution |
Evolutionary algorithm | Genetic programming. | 2D genetic algorithm with problem-specific operators. | Genetic programming training (~25%) and evolution (~75%). | Genetic algorithm. |
Genetic String coding | Four node functions (i.e., and, or, not, and xor) and eight terminal values for pixels around the evolved pixel. | Edge map. Image pixels are masked with corresponding values in pixel map (i.e., 0: no edge, 1: edge). | High-level functions (i.e., avg, min, max, and stdev). Terminal pixels and high-level ephemerals (i.e., gradient and intensity). | Direct bitstream evolution. The solution coding is the actual bitfile. |
Fitness function | Pratt figure of merit (PFM) relative to fault-free Sobel edge detector | Highly complex cost function based on five cost factors. | Biased random sampling fitness evaluation for training. Program fitness is similar to PFM. | Model-free, triplex discrepancy-based function. No application-specific a priori knowledge needed. |
Evolution speed | Partial solution in 2,333 generations after 24 hours of evolution time. | 2,300 generations used for ring imaging; 300 generations used for thin, well-localized edges. | 75 generations, with 25% of images used for training. Very large population size of 2,000. | 148 generations, with low population size of 10. Evolved 8 critical LUTs. |
Best fitness | Not reported | 0.85 PFM with scaling factor of 0.01. | 0.590 for Image 1;0.633 for Image 2. | 100% as compared to output from fault-free Sobel edge detector. |
|