Graph-Based Visual-Semantic Entanglement Network for Zero-Shot Image Recognition

Zero-shot learning uses semantic attributes to connect the search space of unseen objects. In recent years, although deep convolutional network brings powerful visual modeling capabilities ZSL task, its features have severe pattern inertia and lack representation relationships, which leads bias ambiguity. response this, we propose Graph-based Visual-Semantic Entanglement Network conduct graph features, is mapped by using a knowledge graph, it contains several novel designs: 1. establishes multi-path entangled with neural (CNN) (GCN), input from CNN GCN model implicit relations, then feedback modeled information features; 2. attribute word vectors as target for GCN, forms self-consistent regression supervise learn more personalized relations; 3. fuses supplements hierarchical visual-semantic refined into embedding. Our method outperforms state-of-the-art approaches on multiple representative datasets: AwA2, CUB, SUN promoting linkage modelling features.