Image Coding Based on the Automatic Generation of Structural Shapes Using Evolutionary Methods

This study examines the efficient representation of natural objects with structural shapes, such as ferns and snow crystals. Based on the L-System, a formal grammar suitable for representing such structural shapes, evolutionary methods generate rules and parameters to produce images similar to a given image. MSE and SSIM have been the standard measures for quantifying similarity, but they were ineffective for this purpose because they were sensitive to differences in shape. Therefore, in this study, we used the LPIPS rating scale based on deep learning to quantify similarity robust to shape differences. Experimental results confirmed that the codes of the L-System rule that produce images similar to the input image could be obtained.