Multi-level Relationship Construction And Its Application Research In Multi-label Image Classification

The increasing convenience of image acquisition and dissemination has accelerat-ed the growth of the number of images,and the complexity of image content has also increased day by day.Multi-label image classification(MLIC)describes the content of an image with multiple class labels,which can be used in image retrieval,autonomous driving and other fields.The diversity of combinations of multiple labels and image regions and the complexity of the relationship between labels make multi-label image classification still very challenging.In order to construct semantic associations of image regions and labels,and capture the fine-grained correlation between labels,this paper conducts in-depth research of multi-level relationship construction from the perspectives of label maps and image regions.The specific research work is as follows:1)This paper proposes a multi-label image classification method based on Label Hierarchical Graph Inference(LHGI).First,LHGI uses Res Net101 as the backbone network to obtain image visual feature representations.Second,the visual-semantic module uses the classification results to generate class attention maps,which corresponds the label semantics to the image regions,and aggregates the image region features into the label representation to encode the label nodes.Then,the hierarchical graph constructing module generates a fully connected global label graph and dynami-cally generates a local label subgraph according to the importance of the labels.The global label graph and the local label subgraph constitute the hierarchical label graph.Finally,a graph inferencing module based on graph convolution or Transform-er performs multi-level relational inferencing on the hierarchical label graph.LHGI can establish the association between image regions and label semantics,dynamically capture multiple interactions between labels,reduce the redundancy of label relationships,and assist multi-label classification.2)This paper proposes a multi-label image classification method based on Region Relationship Learning(RRL).First,RRL introduces macro-label relationships from datasets in the graph convolution module to achieve global image semantic relationship learning.Second,guided by global information,the local localization module refines the class attention map with image semantics and locates multiple label regions through this attention map.Then,the multi-head attention-based region association module realizes the semantic relation learning of local detail regions and obtains semantically enhanced local feature representations.Then,the multi-head attention-based region association module realizes the semantic relation learning of local detail regions and obtains semantically enhanced local feature representations.This global and local region semantic inferencing can make full use of label relationships,focus on more detailed image features,and then assist label prediction.This paper evaluates our proposed methods from different perspectives,and compares with some mainstream methods on two popular datasets.These datasets have different sizes and different numbers of labels.The quantitative results of the experiments show that our methods outperforms the state-of-the-art methods.