基于感知偏序模型的图标视觉复杂度研究

Study on logo visual complexity assessment based on perceptional relationship model

视觉复杂度分析是计算机视觉研究的重要分支。当前主流研究方法采用基于用户数据的统计概率模型进行定量评估,这样虽能获得统计型结论,但由于缺少对潜在逻辑规则的考虑而无法逼近用户真实模型。用户感知评价的不稳定性常常导致训练模型收敛困难或者次最优现象的出现。本文针对此问题,结合用户感知评价特征与视觉特征的关系,提出基于用户感知偏序关系的视觉图标复杂度分析模型。针对感知数据难以获取和表示的困难,本文提出基于二比较的偏序关系表示用户感知特征;采用可信度预处理减少用户评价数据矛盾冲突对于模型预测的影响。通过提取特定的可视化特征,本文提出改进的SVM模型对基于偏序对的感知数据进行训练获得图标视觉复杂度感知模型。通过进行Pearson、Kendall和Spearman系数的对比,本文预测模型在中国大学图标数据库上与人工评价结果高度相似(>90%)。与最新算法的对比结果证实了本文算法的有效性和先进性。

Visual complexity analysis is one of the key attributes applied in visual computation.The majority of current methods focus on the assessment which is identical to the statistical data through the quantization of special standards.In this way,the statistical results cannot explain the logical rules which are closer to the real user model.The unstable users′label data will cause the difficulties of model convergence and sub-optimal phenomenon.To solve the problems mentioned above,combined with the relationship between user perception evaluation features and visual features,a visual logo complexity analysis model based on user perception partial order relationship is proposed.In view of the difficulty of obtaining and representing perceptual data,this paper proposes a partial order relationship based on two comparisons to represent user perceptual characteristics.Credibility preprocessing is used to reduce the impact of user evaluation data conflicts on model prediction.By extracting specific visualization features,this paper proposes an improved SVM model to train the perception data based on partial order pairs to obtain the icon visual complexity perception model.By comparing the Pearson,Kendall and Spearman coefficients,the prediction model in this paper is highly similar to the manual evaluation results(>90%)in the Chinese University Icon Database.Comparative experiments with the latest algorithms confirm the effectiveness and advancement of the proposed results.