基于前后景分割的图像情感分析  

Image Sentiment Analysis Based on Foreground and Background Segmentation

作　　者：高玮军[1] 刘书君 孙子博 GAO Weijun;LIU Shujun;SUN Zibo(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730000,China)

机构地区：[1]兰州理工大学计算机与通信学院,兰州730000

出　　处：《计算机工程与应用》2025年第1期206-213,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(51668043)。

摘　　要：图像是生活中重要的信息源之一,对其所表达的内容进行细节分析,可以更充分地利用信息资源。随着信息化的快速发展,针对图像模态开展情感分析工作已成为目前研究的一大热点。图像情感分析的主要环节依次为:情感特征提取、情感空间的选择、特征融合和情感识别分类。现有的大部分图像情感分析工作以图像整体为单位进行输入,未能充分发挥图像中局部特征的情感作用。如果不能对图像的全局特征和局部特征作出区分,当图像出现清晰度不高、背景噪声较多等问题时,图像的全局特征就会变得较为敏感,特征提取和识别工作将会受到严重干扰,对情感分析的准确性产生一定影响。针对目前图像情感分析存在的不足,提出一种基于前后景分割的图像情感分析方法。该方法以YOLOv5为框架,引入ConvNeXt模块和AFF模块,分别进行特征提取和注意力融合。实验结果表明,与目前比较流行的几种图像情感分析方法相比,该方法对于包含更多情感信息和语义信息的场景更为适用,性能也有所提升。Images are one of the important sources of information in daily life.By analyzing the details of their expressed content,information resources can be more fully utilized.With the rapid development of information technology,conducting emotional analysis work on image modalities has become a major research hotspot.The main steps of image sentiment analysis are:emotion feature extraction,emotion space selection,feature fusion,and emotion recognition classification.Most of the existing image sentiment analysis work inputs based on the overall image,which fails to fully leverage the emotional role of local features in the image.If the global and local features of an image cannot be distinguished,the global features of the image will become more sensitive when problems such as low clarity and high background noise occur.Feature extraction and recognition work will be severely disrupted,which will have a certain impact on the accuracy of sentiment analysis.In response to the shortcomings of current image sentiment analysis,this article proposes a method for image sentiment analysis based on foreground and background segmentation.This method uses YOLOv5 as the framework and introduces ConvNeXt module and AFF module for feature extraction and attention fusion,respectively.The experimental results show that compared with several popular image sentiment analysis methods,this method is more suitable for scenes containing more emotional and semantic information,and its performance has also been improved.

关 键 词：图像情感分析 前后景分割 特征融合 YOLOv5 局部特征 全局特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]