CBIR中基于最佳路径森林的学习方法

Learning method using optimum-path forest in CBIR

针对高级用户描述的对象与低级图像特征之间的语义差异问题,提出一种基于最佳路径森林的学习方法。使用Gabor小波和鲁棒局部二值模式对查询图像以及待查询数据库图像进行特征提取,获得直方图特征向量;通过最佳路径森林获得图像的相关性反馈,生成标记数据集;通过相似距离度量获知用户的偏好,得到满意结果。实验在Corel数据库和Brodatz数据库上进行,在无噪声的理想情况下,对某些类别的图像,该算法的检索精度高达98%;有噪声情况下,随着噪声的增加,其性能衰减最慢。在Brodatz数据库中,该方法平均检索率只减少了10%左右,在Corel数据库中,较大噪声情况下,其精度比其它方法高出8%左右。

As the issue of semantic differences between objects described by advanced users and low-level description of the image feature, a learning method using optimum-path forest was proposed. The features of the query image and the images to be que-ried in the database were extracted using Gabor wavelet and robust local binary patterns to obtain histogram feature vectors. The relevance feedback of images was obtained using optimum-path forest, generating tags training set. The preference of the users was learnt by similar distance measuring, thus the satisfactory results were got. Results of simulation on Corel database and Brodatz database show, in the ideal case without noise, the retrieval accuracy of some kind of images is up to 98%. Under noisy conditions, the decrease of the performance is the slowest. In Brodatz database, the average retrieval rate is only about 10% re-duction. In the Corel database, the accuracy of the proposed algorithm is higher than that of other outstanding ones by about 8% in the case of big noise.