Generalized Zero-shot Learning Based On Deep Learning

In recent years,deep learning has achieved great success in computer vision and machine learning tasks.However,data-driven supervised learning algorithms rely on huge amounts of labeled data,and are powerless in scenarios where labeled data is scarce or even missing.Zero-shot learning methods can effectively solve many practical problems in the case of limited data.It has important practical significance and has attracted widespread attention.In the absence of unseen training data,the zero-shot learning algorithm uses the semantic knowledge shared by the seen and unseen classes to establish the connection between the visual space and the semantic space,so as to realize the recognition of the unseen classes.However,semantic space and visual embedding space are two completely different spaces and their manifolds are inconsistent,the existing algorithms rely too much on the cross-modal embedding from semantic space to visual space to realize knowledge transfer,and the embedding process lacks effective constraints.In addition,the original semantic attributes,which are highly overlapped for different classes,contain a lot of redundant and invalid information,and even overwhelm the discrimination information between classes,which leads to poorly performance of this kind of methods.Aiming at thses problems,we started from two aspects of feature distribution and attribute learning,and proposed new solutions.The main contributions of the paper can be concluded as:(1)Cross-modal Distribution Alignment Embedding Network(CDAEN)adopts a novel se-mantic embedding network to learn the difference information between classes,so as to enhance the discrimination of semantic embedding features.And CDAEN proposes a distri-bution alignment constraint and an auxiliary classification network to ensure the distribution of semantic embeddings keep consistent with the distribution of real image features,which can alleviate manifolds inconsistent problem.(2)To remove redundant and invalid information in semantic attributes,a Salient Attributes Learning Network(SALN)is proposed to generate discriminative and expressive semantic attributes under the supervision of the multi-label classification results,and l₁₂-is added to enhance the distinguishability of semantic attributes.Then feature alignment network projects salient attributes into visual space and a relation network is adopted for classification.The algorithms proposed in the paper have made significant progress on the five benchmark datasets.The rich and in-depth experiments have verified the effectiveness and excellence of the algorithms.