Presentation on theme: "Interactive Evolutionary Computation Review of Applications Praminda Caleb-Solly Intelligent Computer Systems Centre University of the West of England."— Presentation transcript:
Interactive Evolutionary Computation Review of Applications Praminda Caleb-Solly Intelligent Computer Systems Centre University of the West of England
Summary of Talk Application Areas –Motivation –Implementation –Salient Features Problematic Issues of IEC
Hearing Aid Fitting H. Takagi and M. Osaki Motivation –personalisation of hearing aid compensation characteristics in different acoustic environments Implementation
Hearing Aid Fitting Salient Aspects –Redefinition of filter characteristics based on Gaussian Functions. –Subjective Evaluation of 20 individuals graded on a 5 level scale. –psychological tests to compare clarity and quality using processed sounds from IEC fitting, conventional loudness compensation and unprocessed original sounds.
Image Retrieval 1. H. Takagi, S.B.Cho and S. Noda 2. J.Y.Lee and S.B. Cho 3. F. Boschetti and S.B.Cho Motivation –Enable retrieval of images based on content rather than descriptive keywords allowing incorporation of human preference and emotion search for a specific feature inside an image
Salient Aspects –Correspondence between psychological space and feature space –Evaluation of retrieval performance –Evaluation of features describing content
Image Enhancement R. Poli and S. Cagnoni Motivation –Expertise and knowledge of user required to determine significant regions of interest in images. Implementation –Enhancement of MRI Images –Each program in the population is a solution for altering pixels in the input images to obtain an output image –User drives GP by deciding which individual should be the winner in tournament selection.
Image Enhancement Salient Aspects –Limited user interaction –Modelling the user –Evolutionary algorithms transformed from inefficient search procedures into powerful and efficient search methods.
Problematic Issues of IEC User Fatigue Limited population Limited generations Convergence issues Robustness issues Evaluating Performance
Adaptive Image Segmentation Based on Interactive Evolutionary Search Praminda Caleb-Solly Intelligent Computer Systems Centre University of the West of England
Summary of Talk Description of Application Area Image Processing Technique Interactive Evolution Description of Implementation Evolutionary Algorithm Results Discussion of Research Issues
Components of the Decision Support System Segmentation Feature Extraction Classification Image Capture
Image Segmentation Texture Based Segmentation The Texture Measure Kernel Dimensions Step Size Orientation Angle Threshold Normalise Image Calculate Texture Normalise and Median Filter Threshold and Pad OR Combined Image
Original ImageTexture Image Thresholded Image Segmented Image
Standard Approaches Variety of classical search techniques such as adaptive thresholding and gradient descent used to develop bronze standard set. Process is time and knowledge intensive Not practical for real-time industrial use
Interactive Evolution Three methods for manual intervention by the user –Subjective Selection (Dawkins - Biomorphs) –Subjective Problem Definition (Parmee - Evolutionary Design Systems) –Subjective Evaluation
Description of Implementation 8 IP parameter sets generated at random Parent is the highest scoring individual. 8 offspring produced based on fitness score of parent. User selects new image to score User shown a set of 8 segmented images derived using each of the parameter sets. Images from training set. Calculate aggregate score for each of the 8 parameter sets Best Score > Target Score Yes User sets target score User scores each segmented image on a scale of 0 to 10 Write Results to Log file - Final Parameter sets and corresponding scores.
Evolutionary Strategy (μ,λ) Strategy - (1,8) For Threshold Variables –Mutation Step size depends on the parents fitness For Texture Measures –Depending on the parents fitness the parents texture measure is retained in 50% of the offspring