基于样本不平衡与视觉多样性的超平面偏移法

在基于内容的图像检索中，支持向量机（SVM）的分类性能不仅受到样本不平衡的影响，而且由于图像的视觉多样性，导致在分类超平面附近找不到正例样本，无法提高分类器性能。针对上述问题，提出一种二阶段的SVM超平面偏移方法。根据样本的不平衡性进行超平面的相对偏移，使得当前超平面向理论的最优超平面移动，以此为基础进行相关反馈，并根据反馈结果运用超平面三原则对当前的偏移超平面再进行偏移，以解决图像的视觉多样性问题，从而得到能够提高检索精度的分类超平面。实验结果证明，与基于SVM的标准图像检索方法相比，该方法能大幅提升样本集的分类性能，使图像的检索精度平均提高16％。

Offset Method of Hyperplane Based on Sample Imbalance and Visual Diversity

In the content-based image retrieval, aiming to the problem that the classification performance of Support Vector Machine （SVM） not only is affected by the sample imbalance, but also the visual diversity of images causes positive samples can not be found near the classification hyperplane, and the classification performance can not be improved, this paper proposes an offset method of two stage SVM hyperplane. According to the sample imbalance, the method moves the hyperplane to theoretical optimal hyperplane, and does relevant feedback based on this hyperplane, and according to the result of feedback, it utilizes the three principles of hyperplane to offset the current hyperplane and solves the visual diversity problem, so the better classification hyperplane can be got which has better retrieval precision. Experimental results show that compared with the standard SVM image retrieval method, the method can greatly improve the classification performance of the samples, and has an average of 16% of the performance improvement on retrieval accuracy of image retrieval.