一种基于遗传规划的多特征图像排序算法

A GENETIC PROGRAMMING-BASED MULTI-FEATURE IMAGES RANKING ALGORITHM

提出一种多特征图像的排序算法,通过遗传规划算法对多特征图像排序问题进行建模。利用排序树将候选排序函数表示为遗传规划种群中的个体,把遗传规划算法中的个体由图像特征、常数以及算数运算符组成的三元组构成。随机选择一组初始排序函数作为遗传规划算法的个体,并将其表示成排序树。在每次迭代过程中,计算每棵排序树的适应度,并记录适应度高的排序树,通过变异、交叉以及繁殖操作生成新的个体。交叉操作可以提高种群的多样性,从而实现图像的多样化排序。接下来,从记录过的排序树中,选择对图像训练集排序准确性最高的排序树作为图像排序函数。实验结果表明,所提出的算法能够有效地对图像的多种特征进行融合,在兼顾多样性的同时显著提高了图像排序结果的准确性。

In this paper we propose a multi-feature images ranking algorithm, and model the multi-feature images ranking problem by using genetic programming algorithm. We utilise the ranking trees to represent the candidate ranking functions as the individuals in genetic programming populations, and the individual in genetic programming algorithm of this paper is made up of a triple which contains the image features, constants, and arithmetic operators. A group of initial ranking functions are randomly selected as the individuals in genetic programming algorithm, and then to be organised as ranking trees. In every iteration process, the fitness of each ranking tree is calculated, and then the ranking trees with high fitness value are recorded. Next, new individuals are generated by the mutation, crossover and reproduction operations. Particularly, diversity images ranking can be implemented by the crossover operation which can obviously promote the population diversity. Afterwards, the ranking tree which can rank the images in training dataset with highest accuracy is chosen as the image ranking function. Experimental results show that the proposed algorithm can effectively fuse the multi-features of images, and can significantly improve the accuracy of image ranking results with high diversity.