Development of advanced evolutionary computation methods for complex structure and their applications to motion picture processing

研究了复杂结构的进化计算新算法，包括具有多个数值参数的树形结构、网络结构、图结构等。传统的遗传算法（GA）和遗传规划（GP）等进化算法无法优化这种复杂的结构。我们为它们开发了三种算法。首先，我们开发了一种具有多个数值参数的树形结构优化算法“GMA （Genetic Matrix algorithm）”，并将其应用于图像处理方法的自动构建。其次，构建了以图像滤波器为节点的遗传图像网络GIN （Genetic Image Network）来优化网络结构，并证明了GIN可以自动生成复杂的图像处理程序。第三，我们开发了用于自动编程的GRAPE（图结构程序进化）。GRAPE可以基于学习数据的机器学习生成通用程序。例如，GRAPE可以从几个原始数据及其正确排序的数据的示例中编写排序程序。我们将上述方法应用于图像处理算法的自动生成，特别是运动图像处理，如从运动图像中提取人体和汽车区域。利用进化算法，我们可以很容易地生成复杂的图像处理算法。这些方法可以应用于各种领域。

We studied to develop new evolutionary computation algorithms for complex structures including tree structures with several numerical parameters, network structures, graph structures and so on. Conventional evolutionary algorithms such as GA (Genetic Algorithm) and GP (Genetic Programming) cannot optimize such complex structures.We developed three kinds algorithms for them.First, we developed an algorithm named "GMA (Genetic Matrix Algorithm)" for optimization of tree structures with several numerical parameters and applied it to automatic construction of image processing methods.Second, we constructed GIN (Genetic Image Network) for optimization of network structures whose node is an image filter, and we showed that GIN can generate automatically complex image processing procedures.Third, we developed GRAPE (GRAph structured Program Evolution) for automatic programming. GRAPE can generate general programs based on machine learning of learning data. For instance, GRAPE can make sorting programs from examples of several original data and their correctly sorted data.We applied the above methods to automatic generation of image processing algorithms especially motion picture processing such as extraction of human and automobile regions form motion pictures. We can generate complex image processing algorithms very easily using the developed evolutionary algorithms. These methods can be applied to various kinds of fields.