论文部分内容阅读
[目的 /意义]提出利用社会标签自动分类图片情感类型的方法,服务基于情感特征的图像检索与利用。[方法/过程]以Flickr图片为例,利用PMI算法对Word Net-Affect词表进行预处理形成典型情感词表;结合Ekman提出的6类基本情感类型,利用标签对图片情感类型进行标注;并且,通过实验对分类标注效果进行验证;最后,讨论图片特点、标注意图、非情感标签数量对分类标注效果的影响。[结果 /结论]研究发现,一幅图片的非情感标签与情感标签在表现图片整体情感类型的倾向性上具有较高一致性;结合PMI算法,利用预处理后的典型情感词表标注图片的结果优于未处理的Word Net-Affect词表;并且,分类标注效果与人工标注结果也具有较好的一致性,其中,快乐类(Happy)和忧伤类(Sad)图片的分类标注一致性最高,惊讶类(Surprise)的分类标注一致性最低;分析发现,仅通过标签标注图片情感类型的过程中,分类标注效果与图片情感的典型性、单一性以及图片发布方和欣赏者意图、动机的差异、图片的非情感标签个数都有关系。
[Purpose / Significance] This paper proposes a method of using social tags to automatically classify the emotional types of pictures, and serves the image retrieval and utilization based on emotional features. [Methods / Procedures] Taking Flickr pictures as an example, the Word Net-Affect vocabularies are preprocessed by PMI algorithm to form typical emotion vocabularies. According to the 6 types of basic emotion types proposed by Ekman, labels are used to mark the emotional types of pictures. , And verify the effect of the classification and labeling by experiments. Finally, we discuss the influence of the features of the images, the annotation of the annotations and the number of non-affective tags on the classification and annotation effects. [Results / Conclusions] The study found that the non-emotional tags and emotional tags of a picture have a higher consistency in representing the overall emotional type of the picture. Using the PMI algorithm, the pre-processed typical emotional vocabulary is used to mark the pictures The result is better than the untouched word Net-Affect vocabularies. Moreover, the classification labeling effect is also in good agreement with the artificial labeling results. Among them, Happy and Sad images have the highest consistency of labeling , Surprise (Surprise) classification of the lowest consistency; analysis found that tagging only by the emotional picture of the type of process, the classification of the typical picture of the emotional effects and unity, as well as the image of the publisher and the viewer’s intention, motivation Differences, the picture of the number of non-emotional labels are related.