基于未确知测度的视频情感语义识别

Video affective semantic recognition based on unascertained measure

基于未确知测度理论,提出了一种研究视频情感内容的新算法,建立起视频高层语义与底层特征间的联系.首先,选取能反映视频场景情感变化的视觉类与音频类底层特征,并藉此构建了视频场景对象的视觉类情感特征向量与音频类情感特征向量.其次,分析视频场景情感内容度量的对象空间、指标空间与情感判定空间,确立了视觉、音频情感特征指标的未确测度函数,并建立相应的未确知测度矩阵.最后,采用信息熵权法对各底层情感指标赋值,根据置信度识别准则对视频场景的情感类型进行识别、判定.实验结果验证了该方法的可行性和有效性.

In this paper,a novel algorithm based on the unascertained measure theory is proposed to study the affective contents which are usually contained in video scenes.In order to implement the semantic recognition,several models have been established between the low-level features and the high-level cognitive emotion in video scene. Firstly,on the one hand the low-level visual features named as scene light,shot cut rate and color energy are specially selected as the measure parameters to construct the visual emotion feature vector.On the other hand,five audio com-prehensive features are eventually obtained by dimension reduce,and the auditory emotion feature vector is created accordingly.The visual and auditory features in video scenes are highlighted because of their capabilities to distinguish video affective semantic.Secondly,the unascertained obj ect space is moderately built to formulize each video scene.After that,the emotion index space is consequentially constructed so that a series of unascertained measure functions are respectively formed to quantify the each low-level feature.Furthermore,the emotional decision space is reasonably designed to bridge the gap between the unascertained obj ect space and the emotion index space. Then,the specific process of building the measure matrix is discussed by means of the unascertained measure theory in detail.Finally,the information entropy is adopted to determine the coefficients of weight in the unascertained measurement models.According to the results from above analysis,the degree of confidence concept is introduced to evaluate which emotion type could exactly match the affective content in the target video scene.A series of thenbsp;contrast experiments were carefully performed to test the robust performance of the method.The experimental results verify the feasibility and effectiveness of the proposed algorithm.