Abstract: In allusion to that the single feature and manual setting multi-dimensional weighting features cannot satisfy the need of CBIR more and more, we presents a new algorithm of multi-dimensional features vector weighting based on the cluster of training sample. The algorithm needs to manually build the training sample, extract the multi-dimensional features like color, texture and shape of each element, and then use the genetic algorithm to find the most optimal weighting coefficient set of features vectors sample. Finally, it calculates the proper value of the training sample using the set and retrieve the sample. Proved by the experiment, this algorithm can improve the classified accuracy compared to other algorithm and has a high accuracy in differentiating two clusters with high similarity.

Key words: image retrieval, training sample cluster, multi-dimensional features vector, GA, most optimal weighting coefficient set

摘要： 单一特征检索图像和手工设置多维加权系数特征检索图像越来越不能满足基于内容图像检索精度的需要,为此提出一种基于训练样本集聚类的多维特征向量加权算法。该算法需要手工建立训练样本集，提取出每个图像的颜色、纹理和形状等多维特征，使用遗传算法寻找特征向量集的最优加权系数序列，最后使用该加权序列计算测试集的特征值进行图像检索。实验证明，该算法相对于单一特征检索和手工设置多维特征加权在检索的准确度上有一定的提高，并且在相似度比较高的两个聚类检索时，有很高的准确性。

关键词: 图像检索, 训练样本集, 多维特征向量, 遗传算法, 最优加权系数