Auxiliary objectives for the evolutionary multi-objective principal color extraction from logo images
Auxiliary objectives for the evolutionary multi-objective principal color extraction from logo images
复制标题
从标志图像中进化多目标主色提取的辅助目标
DOI：
10.1109/cec.2008.4631276
复制
发表时间：
2008
期刊：
影响因子：
--
通讯作者：
Kaori Yoshida
中科院分区：
文献类型：
--
作者：
M. Köppen;Yutaka Kinoshita;Kaori Yoshida
在本文中，我们提出了一种为徽标图像类别选择主要颜色的方法。该方法使用可分配给颜色集的多个目标，将所选颜色限定为图像的主要颜色。由于所有这些目标都有不同的偏好，并且具有不同的计算复杂性和粒度，因此将它们全部放入单个目标向量中是没有用的。相反，建议采用三阶段程序。第一阶段仅优化高相关性和较低计算量的目标。这里，使用进化多目标优化。第二阶段根据一组附加目标重新评估第一阶段的帕累托集。最后，根据最高偏好的单个目标选择第二阶段帕累托集的一个解决方案。作为第一阶段的合适目标，已经找到了颜色集到图像像素的平均最小距离，以及比阈值更近的像素的平均数量。该方法在许多标志图像上进行了研究，在大多数情况下，它可以根据找到的主色重建具有良好视觉质量的标志图像。实验还表明，通过搜索比徽标图像中存在的更多主色，并重复该过程以找到小但值得注意的细节结构，通常可以改善结果。
In this paper, we present an approach to the selection of principal colors for the class of logo images. The approach is using multiple objectives that can be assigned to a color set, qualifying the selected colors as being principal colors of the image. Since all these objectives have a different preference, and have different computational complexity and granularity, it is not useful to put them all together into a single objective vector. Instead, a three stages procedure is proposed. The first stage optimizes only objectives of high relevance, and lower computational effort. Here, evolutionary multi-objective optimization is used. The second stage re-evaluates the Pareto set of the first stage according to an additional set of objectives. Finally, one solution of the Pareto set from the second stage is selected according to a single objective of highest preference. As suitable objectives for the first stage, the average minimum distance of the color set to the image pixels, together with the average number of pixel that are closer than a threshold have been found. The approach was studied on a number of logo images, and it could reconstruct the logo images of good visual quality from the found principal colors in the majority of the cases. The experiments also show that the result is usually improved by searching for more principal colors than are present in the logo image, and by repeating the process to find also small, but notable detail structures.