Painting Using Genetic Algorithm with Aesthetic Evaluation of Visual Quality - Technologies and Applications of Artificial Intelligence - page 130

This study adopts the well-known uniform crossover for the GA. Considering
the fact that the two parents may have different numbers of agents, the crossover
operation is performed with the smaller number of agents between the two par-
ents. Then the uniform crossover sequentially determines at random which par-
ent for a gene to inherit from. The two offspring thus hold the characteristics of
both parents.
For the mutation, we adopt the uniform mutation. This mutation scans the
genes of offspring and decides to change a gene to a random value according to
the mutation rate. For the genes representing agents, if mutation occurs, it will
randomly generate a new agent to replace the old one. This mutation is expected
to increase the diversity.
As for survival selection, this study utilizes the ( μ + ʻ ) strategy to keep elitists
among the population.

3.3 Fitness Function

In this study, we propose using two measures of visual quality, i.e., edge distri-
bution and line intersection, in place of human feedback to evaluate the quality
of painting.

Edge Distribution
In evaluating the fitness of a painting, we first consider that the important
regions should have more focus points. Li et al. [8] have validated the utility of
this feature. Given that our painting comprises the results from several agents
coloring the canvas pixel by pixel and thus has interactions, we remove the
Gaussian smoothing filtering used in the original method and keep the edge
noises to increase edge points in the generated paintings.
Specifically, we convert the image from HSB scheme to gray channels and then
apply a 3 3 Laplacian filter on it (cf. Fig. 4). Afterward, we calculate the area

of the smallest bounding box and its ratio to the whole area. The bounding box
is defined as the rectangle that contains 81% edge points [8]. Figure 5 shows the
original image and the image after applied with Laplacian filter and its bounding
box. According to the test in [8], high-quality paintings have a low ratio, whereas
low-quality paintings have a high ratio. Therefore, we define the complement of

×

Fig. 4. Discrete approximation to the Laplacian filter