CAJ | 학술논문

大部分传统的图像自动标注方法需要训练数据中具有精准的标注词，然而这样的数据通常是需要人工标注的，因此获取成本较高，且存在一定的主观性。该文提出一种全新的图像自动标注方法，通过结合自然语言理解领域实体识别的技术，充分利用图像周边自带环绕文本，将图像视觉特征、环绕文本以及实体抽取所得到的能够描述图像中显著特征的词在概率主题模型中进行联合建模，学习到多种数据模态之间的关联关系，从而实现图像的自动标注。在UIUC Pascal Sentence数据集上的实验证明该方法比传统方法具有更好的图像标注预测以及检索性能。
대부분전통적도상자동표주방법수요훈련수거중구유정준적표주사，연이저양적수거통상시수요인공표주적，인차획취성본교고，차존재일정적주관성。해문제출일충전신적도상자동표주방법，통과결합자연어언리해영역실체식별적기술，충분이용도상주변자대배요문본，장도상시각특정、배요문본이급실체추취소득도적능구묘술도상중현저특정적사재개솔주제모형중진행연합건모，학습도다충수거모태지간적관련관계，종이실현도상적자동표주。재UIUC Pascal Sentence수거집상적실험증명해방법비전통방법구유경호적도상표주예측이급검색성능。
Most of the traditional approaches for automatic image annotation generally demand on training data with exact labels .However ,this kind of data usually comes from human generated tags ,which are too subjective and often difficult to obtain .In this paper ,we propose a novel image annotation model which can utilize the rich surrounding text of images by integrating entity extraction technology of Natural Language Processing .Then the image features ,surrounding text and extracted entity words which we assume that can more directly describe the salient objects in the corresponding ima‐ges are modeled in a uniform probabilistic topic model framework .The learned correlations among different data modali‐ties can be used in image annotation afterwards .Experimental results on UIUC Pascal Sentence dataset show that our model outperforms the traditional topic model‐based image annotation methods in both annotation and retrieval .