基于典型相关子空间与K邻近的图像自动注释算法

Image Automatic Annotation Algorithm Based on Canonical Correlation Analytical Subspace and K-Nearest Neighbor

本文提出了一种图像自动注释算法.为了建立图像视觉特征和语义特征间的联系,提出利用典型相关方法将这两种特征映射到同一特征子空间,进而得到每一个语义标签的图像子集.针对弱标记问题,提出重新定义每个语义标签的语义空间,将典型相关特征子空间的图像子集与训练集中每个标签原有的图像进行融合,形成更加完善的图像子集.针对图像训练集类别不平衡问题,提出在每个标签的语义空间中利用K邻近算法求得每个标签的语义邻居,使每个标签的图像个数相对平衡.在注释过程中,利用每个语义子集与待注释图像的视觉距离建立贝叶斯概率模型完成注释工作.在基准测试图像库进行对比实验,结果表明,所提出的算法能够有效地完成图像自动注释工作.

An automatic image annotation algorithm was proposed.In order to establish the relationship between the visual features and the semantic features of the image,the two features were mapped to the same feature subspace by canonical correlation method,and then the image subsets of each semantic label are obtained.In view of the problem of weak markup,the semantic space of each semantic label was redefined,and the image subsets of the typical correlation feature subspace were fused with the original image of each label in the training center to form a more perfect subset of the image.In view of the problem of the class imbalances of the image training set,the semantic neighbors of each label were obtained by using the K proximity algorithm in the semantic space of each label,making the number of each label relatively balanced.In the annotation process,Bayesian probability model was established using each semantic subset and the visual distance of the image to be annotated to complete the annotation work.The result of the contrast experiment in the base test image library shows that the proposed algorithm can effectively complete the automatic image annotation.