Dynamic Semantic Complementary Network for Zero-Shot Learning

Zero-Shot Learning (ZSL) transfers knowledge from seen to unseen classes by associating visual features with shared semantic information. However, semantic information consists of high-level attributes with complex visual characteristics and the contrasting low-level attributes. Existing ZSL methods tend to oversimplify visual-semantic associations by ignoring high-level attributes, resulting in incomplete utilization of semantic information and limited knowledge transfer. To address this issue, we propose a novel Dynamic Semantic Complementary Network (DSCN). Based on training a main subnet that obtains complete semantic information, DSCN dynamically calculates and selects the ignored semantic information with a novel semantic utilization metric. An auxiliary subnet is then used to learn the ignored semantic information. Finally, we deploy complementary knowledge distillation to conduct effective knowledge interactions between the two subnets. DSCN makes full use of semantic information through the mutual complementarity between two subnets. Extensive experiments on three benchmark datasets (CUB, SUN and AWA2) demonstrate that DSCN achieves superior performance by maximizing semantic utilization, surpassing existing state-of-the-art methods.