基于不平衡数据与集成学习的图像情感分类研究

Research on Image Sentiment Classification Based on Unbalanced Data and Ensemble Learning

目前,基于图像的情感分析已成为情感计算领域的研究热点。针对图像情感分析常用的开放数据集通常表现为多分类的不均衡数据,单一模型存在抽取特征单一、泛化能力不强等问题。首先,改进Focal Loss损失函数,使模型跟随训练进度动态调整聚焦参数。然后,设定概率阈值参数确定困难样本,通过舍弃困难样本避免模型学习错误特征。接下来,选取3个分类性能良好的卷积神经网络模型作为基分类器,分别关注图像的局部、颜色及深度语义特征。最后,采取加权投票法策略,引入信息熵更新多分类器决策的权值。实验表明,所提方法能提升图像情感多分类的准确率,可为基于不平衡数据与集成学习的图像情感分类研究提供参考与借鉴。

At present,image based affective analysis has become a research hotspot in the field of affective computing.The open data set com⁃monly used in image emotion analysis is usually presented as multi category unbalanced data,and the single model has the problems of single feature extraction and weak generalization ability.First,the Focal Loss loss function is improved to make the model dynamically adjust the fo⁃cus parameters following the training progress.Then,probability threshold parameters are set to determine the difficult samples,and the mod⁃el avoids learning incorrect features by discarding the difficult samples.Next,select three convolutional neural network models with good clas⁃sification performance as base classifiers,focusing on the local,color,and depth semantic features of the image.Finally,the weighted voting strategy is adopted,and information entropy is introduced to update the weights of multi classifier decisions.The experiment shows that the pro⁃posed method can improve the accuracy of image sentiment multi classification,and can provide reference for the research of image sentiment classification based on imbalanced data and ensemble learning.