Multilabel Recognition Algorithm With Multigraph Structure

Abstract

In multi-label recognition tasks, convolutional neural networks (CNNs) can extract rich features of images in Euclidean space. However, CNNs struggle to handle dependencies between multi-labels in non-Euclidean space. Graph Convolutional Networks (GCNs) have natural advantages in processing the data in non-Euclidean space. Therefore, many excellent algorithms used GCN to capture the deep dependencies among multi-labels. However, these algorithms directly applied GCN for label feature fusion, leading to nodes with background features affecting the quality of the representation of their neighbor nodes. For example, during label feature smoothing, the representation of nodes with weak edge information tends to converge to nodes with background features. We propose a multi-label image recognition algorithm with a multi-graph structure to address the impact of background features on recognition performance. In brief, when smoothing the dependencies among labels using GCN, we transform the shared adjacency matrix into a combination of multiple adjacency matrices. Unlike node features sharing an adjacency matrix, the multi-graph structure can calculate the performance of each feature carried by the nodes separately during feature fusion and convey more appropriate information rather than in a generalized manner. We have conducted extensive experiments on two multi-label image recognition datasets and have achieved the state-of-the-art.