基于显著点特征多示例学习的图像检索方法

A method for image retrieval based On salient points feature multiple-instance learning

提出了一种基于图像显著点特征进行多示例学习(Multiple-instance learning)的图像检索方法。该方法对图像进行小波分解并跟踪不同尺度小波系数提取图像显著点;然后利用显著点特征进行检索,并在相关反馈中将图像看作多示例包,通过期望最大多样性密度(EM-DD,expectation maximization diverse density)方法进行多示例学习,获得体现图像语义的目标特征。在Corel和SIVAL两个图像库进行实验,结果表明该方法明显提高了检索的准确性。

A novel method based on multiple-instance learning with salient points is presented for content based image retrieval. This method extracts salient points of image by tracking wavelet coefficients of different scales,and then the global statistical feature of salient points is used for image retrieval. In relevant feedback,images containing salient points are treated as multiple-instance bag,and trained using expectation maximization diverse density（EM-DD） algorithm. A target feature representing the image content is obtained in the training and used for retrieval Experimental results from Cord and SIVAL image database illustrate the validity of the proposed framework.