基于词频同现与WordNet的图像自动标注改善算法研究

图像自动标注是模式识别与计算机视觉等领域中的重要问题。针对现有图像自动标注模型普遍受到语义鸿沟问题的影响,提出了基于关键词同现的图像自动标注改善方法,该方法利用数据集中标注词间的关联性来改善图像自动标注的结果。此外,针对上述方法不能反映更广义的人的知识以及易受数据库规模影响等问题,提出了基于语义相似的图像自动标注改善方法,通过引入具有大量词汇、包含了人知识的结构化电子词典WordNet来计算词汇间的关系并改善图像自动标注结果。实验结果表明,提出的两个图像自动标注改善方法在各项评价指标上相比以往模型均有所提高。

Automatic image annotation refinement based on keyword co-occurrence and WordNet

Image automatic annotation is a significant and challenging problem in pattern recognition and computer vision areas. At present, most existing image annotation models are influenced by semantic gap problem. This paper proposed a new image automatic annotation refinement method based on keyword co-occurrence to overcome above problem, which used the correlations between keywords in dataset to improve image annotation result. However, above method did not reflect the generalized knowledge of people and easy influenced by the size of dataset. Aiming at above problem, it proposed a new image automatic annotation refinement method based on semantic similarity to overcome above problem. This method used semantic dictionary WordNet to calculate the correlations between keywords and improve the image annotation results. Experimental results conduct on Corel 5K datasets verify the effectiveness of proposed image annotation method. The proposed automatic image annotation model improves the annotation results on all evaluation methods.