Evolvable Image Filter Design With Mixed Constraints

This research is based on laboratory’s project. This paper elaborates the way howwe evolve the original Genetic Algorithms(GA)，and evolve the image filter with theevolved GA. At present, the fact that customer requests much more higher qualityimage，the image processing technology should be advanced with request. The evolvedGA we propsed is able to generat more optimized image filter to satisfy the request.Evolutionary Hardware (EHW) was proposed to use simulated evolution inimprovement of the circuit, including both the layout and the performance. EvolutionaryAlgorithms (EA) was one of the algorithms used in EHW, which made use of stochasticsearch algorithms to find out more optimized circuits. GA was one method belonging toEA. It was firstly proposed to model adaptation of natural and artificial systems throughevolution. By keeping a population of chromosomes which representing possiblesolutions, evolution process was executed with the aim of reaching an optimal solution.Image filter was used as a preprocessing unit before recognition, as imagesacquired by modern cameras were often contaminated with different noise types.Recently several evolutionary approaches have been used to deal with circuit structuresof image filters, including GA as before mentioned, due to the needs of intelligentprocessing. Some objectives concerning complexity signal delay and powerconsumption of the circuit layout were also proposed. However, some problems arosesuch as processing time, image quality and scalability issue.In this research, some proposals are raised in evolution process by fixing whole orpart of the chromosome representation of the image filter, together with other geneticoperators one-point mutation and deterministic selection. It aims to find out bettercircuit structure by using circuits from previous experiments with best fitness valuesthrough mutation and crossover. Therefore, the Elite Fitness values and MeanDifference Per Pixel (MDPP) values should be improved than earlier methods. They arecarried out in processes as:(1) chromosome representations of the proposed circuit aredefined, either fixed or partially-fixed;(2) multi-objectives are defined mainly to fitnessfunction F2;(3) performances are evaluated by Elite Fitness values of the circuit andMDPP values of the filtered images.The target of this research is to find out circuit layouts with better Elite Fitnessvalues which produce better MDPP values. The simulation is performed by using Eclipse SDK together with Java Runtime Environment. By using several differentlystructured image filter representations, which refers to chromosomes here, results areobtained that: the Elite Fitness values as well as the MDPP values are all improved byimage filters constructed by proposed methods. Therefore, the performance of imagefilters can be improved as well.