Evolutionary image processing

from Wikipedia, the free encyclopedia

The evolutionary image processing (EB) is a field of digital image processing . Evolutionary algorithms (EA) are used to optimize and solve various image processing problems. Evolutionary image processing thus represents the combination of evolutionary optimization and digital image processing.

EA has been used in computer science for several decades to optimize various problems. The application in image processing, on the other hand, is still a very young research area. This is primarily due to the technological development of computer systems, since EB is a relatively computationally intensive process.

application areas

Reconstruction of image processing algorithms

Problem with the reconstruction of an image processing algorithm

By applying different operations to an image, sequential processing creates an image operation chain . Image processing programs or medical and industrial systems that work with visual systems often contain image processing methods whose functions cannot be fully viewed. If an image is edited, a so-called result image is created next to the original image. A manual reconstruction of the image operation chain, which lies between the original image and the resulting image, is not possible in most cases. In evolutionary image processing, an EA is adapted so that the chain of operations can be automatically reconstructed.

Optimization of image processing algorithms

With the help of the reconstruction it is possible to optimize various image processing problems. Shorter image operation chains can be found by adapting the EA. By finding shorter solution paths, the computing time of the solution method for the problem is reduced. Through a targeted weighting of the image operations, lighter operation chains that require a shorter computing time can also be found.

Adaptation of the evolutionary algorithm

Genes and individuals

Genes of an individual in EB

The adaptation of different individuals is a sufficient criterion for the use of an EA. The information that can be obtained from a filter chain must be mapped to an individual accordingly. A filter can represent a gene. The exact order of the filters that the chain contains is a feature that enables uniqueness. Thus the order of a chain can be transferred to the chromosome of an individual. Each individual can thus be assigned a function in which they work through the sequence according to their genes. The original picture of the problem is available to each individual. A corresponding result image then results from its function, with the help of which the fitness can be determined. The size of an individual can vary from 1 to the maximum specified size. By generating many individuals, the collection can be combined into a population .

Genetic operators

In order to develop and continue partial results, genetic adaptations are necessary. The generated individuals are granted the right to reproduce so that they can pass on their genes. The recombination of certain individuals creates the appropriate subsequent generation, which contains the appropriate genes for the problem.

Recombination

Recombination in EB

A so-called one-point crossover is suitable for the recombination of two individuals . In this crossover, half of the father genes and half of the mother genes are transferred to a new individual with the help of a crossing point. Before recombination, it is first necessary to determine the length of the child's chromosome (length of the child). This happens over the lengths of the chromosomes of the father and mother. Applying a one-point crossover to two parents creates two children (so-called genomes). However, since only one child should result from the recombination, the child is determined at random from the two genomes. The two genomes are inherited from the halves of the respective parent genes. The genomes differ in the combination of the two halves.

mutation

Mutation in EB

With the help of the mutation , diversity within the population can be maintained and increased. The mutation occurs by replacing a filter within the chain with a randomly chosen filter. The mutation is carried out after recombination and only applied to part of the population. With the help of a random variable, a certain number of child individuals is changed by the mutation after the recombination.

Properties of evolutionary image processing

advantages

A reconstruction of image processing algorithms is possible with the help of the EB. The EA systematically continues good partial results and in most cases finds the chain of operations faster than with manual reconstruction. By finding algorithms that are less computationally intensive, different approaches can be optimized.

disadvantage

EA usually have a large computing time requirement. In combination with operators of image processing and depending on the size of the image data, the computing time required is increased. The runtime can be improved with the help of parallel programming . Optimization using graphics processors is also possible.

literature

  • Roman Kalkreuth, Jörg Krone, Michael Schneider: Automatic generation of image operation chains by means of genetic programming , University of Applied Sciences South Westphalia, Institute for Computer Vision & Computational Intelligence, in: Proceedings 22nd Computational Intelligence Workshop , pages 325-340, 2012, ISBN 978-3-86644 -917-6 .
  • Marc Ebener: Evolutionary Image Processing , Gesellschaft für Informatik e. V., 2008.
  • Marc Ebner: Color Constancy , John Wiley & Sons, 2007 ISBN 0-470-05829-3 .
  • Marc Ebner: Evolving color constancy. Special Issue on Evolutionary Computer Vision and Image Understanding of Pattern Recognition Letters , 2006.
  • Christoph Bullmann: Evolutionary Image Processing (PDF), HTWK Leipzig, Department of Mathematics, Computer Science and Natural Sciences, 2009.
  • Jun Ando, ​​Tomoharu Nagao: Fast Evolutionary Image Processing Using Multi-GPUs Yokohama National University, 2009, ISBN 978-953-307-026-1 .

Individual evidence

  1. ^ A b Roman Kalkreuth, Jörg Krone, Michael Schneider: Automatic generation of image operation chains by means of genetic programming, University of Applied Sciences South Westphalia, Institute for Computer Vision & Computational Intelligence, in: Proceedings 22nd Workshop Computational Intelligence, pages 339-340, 2012, ISBN 978- 3-86644-917-6 .
  2. ^ Roman Kalkreuth, Jörg Krone, Michael Schneider: Automatic generation of image operation chains by means of genetic programming , University of Applied Sciences South Westphalia, Institute for Computer Vision & Computational Intelligence, in: Proceedings 22nd Workshop Computational Intelligence , pages 327-328, 2012, ISBN 978-3- 86644-917-6 .
  3. ^ Roman Kalkreuth, Jörg Krone, Michael Schneider: Automatic generation of image operation chains by means of genetic programming , University of Applied Sciences South Westphalia, Institute for Computer Vision & Computational Intelligence, in: Proceedings 22nd Workshop Computational Intelligence , pages 329-330, 2012, ISBN 978-3- 86644-917-6 .
  4. ^ Roman Kalkreuth, Jörg Krone, Michael Schneider: Automatic generation of image operation chains by means of genetic programming , University of Applied Sciences South Westphalia, Institute for Computer Vision & Computational Intelligence, in: Proceedings 22nd Workshop Computational Intelligence , page 330, 2012, ISBN 978-3-86644- 917-6 .
  5. ↑ Jun Ando, ​​Tomoharu Nagao: Fast Evolutionary Image Processing Using Multi-GPUs Yokohama National University, 2009, ISBN 978-953-307-026-1 .

Web links