Design of composite image filters using interactive genetic programming

We combine a method for designing composite image filters with interactive genetic programming (IGP). Human subjective tests are used to comparitively evaluate the IGP-based filter design method to a manual filter design method and the multi-stage filtering feature of a software photo-retouching program. The composite image filter has a tree structure, with its nodes consisting of multiple simple image filters, arithmetic operators, arithmetic functions, constant values, and the pixel value of the input image. Genetic programming (GP) optimizes the tree structure based on the visual inspection of the IGP users, i.e. filter designers. Ten filter designers design composite filters using three methods: an IGP-based design method, a manual-based design method, and using the photo-retouching features of a commercial software program to time-sequentially apply ready-made filters. The designers make filters that output images corresponding to the given design concepts - relaxed and violent - based on their visual inspection. Twenty subjects compare the obtained images in pairs and evaluate which image is closer to achieving the give design concept. Wilcoxon signed-rank test demonstrates that the IGP-base filter design method can produce filters that create images with impressions that are closer to the given design concept than the other two methods.