基于 CNN 和加权贝叶斯的最近邻图像标注方法-《计算机技术与发展》

文章信息/Info

Title:: A Nearest Neighbor Image Annotation Method Based on CNN and Weighted Bayesian

Author(s):: WANG Lin; ZHANG Su-lan* ; YANG Hai-feng; School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China

Keywords:: automatic image annotation; nearest neighbor model; mapping relationship; convolutional neural networks; Bayesian posterior probability

摘要:: 图像标注的准确性在很大程度上关系着图像检索的准确性。然而, 传统的基于最近邻模型的图像自动标注方法不能有效提取图像底层特征, 并且无法有效建立低级视觉特征到高级语义之间的映射关系,使得近邻图像搜索不准确从而影响图像标注的准确性。针对上述问题,提出了一种改进的基于 CNN 和加权贝叶斯的最近邻图像标注方法。首先,利用卷积神经网络( convolutional neural networks,CNN)提取图像特征,并依此特征搜索其近邻图像,构建候选标签集合;然后利用贝叶斯后验概率构建待标注图像的视觉特征与标签之间的映射关系;最后通过设定权重优化概率值并排序,得到最优的候选标签进而实现图像标注。在三个基准数据集 Corel 5K,IAPRTC-12 和 ESP Game 上进行实验,结果表明该方法在准确率、召回率与 F1 值上均取得了较好的效果。

Abstract:: The accuracy of image annotation is related to the accuracy of image retrieval to a great extent. However,the traditional image automatic annotation method based on the nearest neighbor model cannot effectively extract the features of the image,and the mapping relationship between low-level visual features and high - level semantics cannot be effectively established either. So the accuracy of the nearest neighbor images and labels is low. To solve the above problems,a neighbor image annotation method based CNN and weighted Bayesian is proposed. First,image features are extracted by using CNN,from which the nearest neighbor images are searched and the candidate labels are obtained. Then the mapping relationship between visual features and labels of the unlabeled image based on Bayesian posterior probability is constructed. Finally,the probability value is optimized and sorted according to set the weight, from which the optimal candidate labels are selected to realize image annotation. Experiments on three benchmark datasets Corel 5K, IAPRTC-12 and ESP Game show that the proposed method achieves better results in precision,recall and F1 value.