Automated Evaluation and Generation of Graphic Arrangements through Adaptive Evolution

This study explores methods to partially automate a graphic design task through evolutionary computation. Automating visual evaluation is not a simple task, as visual characteristics, style and taste are often hard to define, or to parameterize. Regarding this problem a series of target images are used to define desired formal characteristics. Through these images, target features (describing color distributions and spatial arrangements) for a series of 2D arrangements are determined separately. This separation enables us to create novel images using different targets for each feature type. In this study, three usages for the extracted features are proposed: 1. For the initiation steps of the evolutionary processes, as distribution maps. 2. As 1D histograms for measuring color similarity between the target distribution and candidate arrangements. 3. For grid-based layout evaluation, in order to compare the layouts in the form of rows, columns and cells. An adaptive multi-objective evolutionary process is implemented through a consecutive micro-processes approach, which is suggested to mimic the human generate-and-test procedure. Four objective functions with corresponding mutation operators are utilized at the same run consecutively, according to the course of the process. Several test series are carried out and the approach is found applicable: 1. Each of the four objective functions is tested separately, without initiation maps, for adaptive and non-adaptive versions. While all of them successfully converged to desired levels, non-adaptive versions were slightly better. 2. For 12 scenarios, four objective functions are used together in a multi-objective process setting, with initiation maps. All scenarios succeeded in maintaining and/or obtaining desired fitness levels. Yet, when compared, non-adaptive versions are found more successful and reliable for our parameter and problem settings.