Semantic-guided Reinforced Region Embedding for Generalized Zero-Shot Learning

Generalized zero-shot Learning (GZSL) aims to recognize images from either seen or unseen domain, mainly by learning a joint embedding space associate image features with the corresponding category descriptions. Recent methods have proved that localizing important object regions can effectively bridge semantic-visual gap. However, these are all based on one-off visual localizers, lacking of interpretability and flexibility. In this paper, we propose novel Semantic-guided Reinforced Region Embedding (SR2E) network localize objects in long-term interests construct space. SR2E consists Module (R2M) Semantic Alignment (SAM). First, without annotated bounding box as supervision, R2M encodes semantic guidance into reward punishment criteria teach localizer serialized region searching. Besides, explores different action spaces during searching path avoid local optimal localization, which thereby generates discriminative less redundancy. Second, SAM preserves relationship via alignment designs domain detector alleviate confusion. Experiments four public benchmarks demonstrate proposed is an effective GZSL method reinforced space, obtains averaged 6.1% improvements.