一种融合标签间强相关性的多标签图像分类方法  

A Method of Multi-label Image Classification with Fusing Powerful Semantic Correlation

作　　者：张辉宜 夏媛龙 周克武 包向华 陶陶 ZHANG Huiyi;XIA Yuanlong;ZHOU Kewu;BAO Xianghua;TAO Tao(School of Computer Science and Technology,Anhui University of Technology,Anhui Maanshan 243000,China;Ma’anshan Public Security Bureau,Anhui Maanshan 243000,China)

机构地区：[1]安徽工业大学计算机科学与技术学院,安徽马鞍山243000 [2]马鞍山市公安局,安徽马鞍山243000

出　　处：《重庆工商大学学报（自然科学版）》2023年第5期8-15,共8页Journal of Chongqing Technology and Business University:Natural Science Edition

基　　金：安徽省重点研发计划项目(201904D07020020);安徽省自然科学基金项目(1908085MF212).

摘　　要：为了将标签间的语义相关性引入多标签图像分类模型中,传统的方法例如ML-GCN通过设置单阈值将标签条件概率矩阵二值化为标签共现矩阵,然而,仅设置单阈值很难归纳所有的标签语义关系情况。针对这一问题,提出一种融合标签间强相关性的多标签图像分类方法—MGAN(Multiple Graph Convolutional Attention Networks),通过设置多个阈值,将传统的标签条件概率矩阵按照不同的相关性程度分割为多个子图;同时,为了提升多标签分类性能,也引入图像区域空间相关性。另外,针对传统的“CNN+GCN”方法将标签与特征的融合张量视为预测分数缺乏可解释性问题,将标签与特征的融合张量视为注意力分数;在MS-COCO和PASCAL VOC数据集上与其他主流多标签图像分类方法进行了对比实验,平均准确率分别达到了94.9%和83.7%,相较于经典ML-GCN模型,分别获得了0.9%和0.8%准确率提升,且在“Binary”和“Re-weighted”邻接矩阵模式下,MGAN都有较好的表现,验证了新的融合方法可以缓解图卷积神经网络过平滑问题对多标签图像分类的影响。In order to introduce semantic correlation between labels into multi-label image classification model,traditional methods,such as ML-GCN,transform label conditional probability matrix into label co-occurrence matrix by using single threshold value.However,it is difficult to sum up all semantic relationships of all labels by using single threshold value.To solve this problem,a method of multi-label image classification with fusing powerful semantic correlation,MGAN,was proposed.By setting multiple thresholds,the traditional conditional probability matrix of labels was divided into multiple subgraphs according to different degrees of correlation.Meanwhile,in order to improve the performance of multi-label classification,image region spatial correlation was also introduced.In addition,the traditional“CNN+GCN”method regards the fusion tensor of label and feature as the lack of interpretability of the predicted fraction.To solve this problem,MGAN regards the labels and feature’s fusion tensor as the attention score.Compared with other mainstream multi-label image classification methods on MS-COCO and PASCAL VOC datasets,the mAP were 94.9%and 83.7%respectively,which were 0.9%and 0.8%higher than traditional ML-GCN model.And MGAN performed well in both“Binary”and“Re-weighted”adjacency matrix mode,which verified that the new fusion method can alleviate the influence of graph convolutional neural network’s“over smoothing”problem on multi-label image classification.

关 键 词：多标签图像分类 语义相关性 图卷积网络 注意力机制 区域空间相关性 

分 类 号：O643[理学—物理化学]