场景语义树图像标注方法

Automatic image annotation based on scene semantic trees

自动图像标注是一项具有挑战性的工作,它对于图像分析理解和图像检索都有着重要的意义。在自动图像标注领域,通过对已标注图像集的学习,建立语义概念空间与视觉特征空间之间的关系模型,并用这个模型对未标注的图像集进行标注。由于低高级语义之间错综复杂的对应关系,使目前自动图像标注的精度仍然较低。而在场景约束条件下可以简化标注与视觉特征之间的映射关系,提高自动标注的可靠性。因此提出一种基于场景语义树的图像标注方法。首先对用于学习的标注图像进行自动的语义场景聚类,对每个场景语义类别生成视觉场景空间,然后对每个场景空间建立相应的语义树。对待标注图像,确定其语义类别后,通过相应的场景语义树,获得图像的最终标注。在Corel5K图像集上,获得了优于TM(translation model)、CMRM(cross media relevance model)、CRM(continousspace relevance model)、PLSA-GMM(概率潜在语义分析-高期混合模型)等模型的标注结果。

Automatic image annotation （AIA） is an important and challenging job for image analysis and understanding such as Content-Based Image Retrieval （CBIR） . In AIA, a model for annotation which represents the relationship between the semantic concept space and the visual feature space, is constructed by learning from annotated image datasets. The performance of the AIA is still poor because the relationship between the key words and visual features is too compli- cated due to the semantic gap. However, with constrains under the scenes, the correlation between them becomes simpler and clearer for better annotation results. In this paper, a method of automatic image annotation based on scene semantic tree is proposed. Image scenes are obtained through the annotated words using PLSA. Scene semantic trees are constructed for each image scene. Un-annotated images are classified into certain scenes with visual features and are annotated using the corresponding scene semantic tree. Using the visual features provided by Duygulu, the experiments get the more effective results on Corel5K database than TM, CMRM, CRM and PLSA-GMM.