Research On Zero-shot Learning Algorithm Based On Feature Embedding And Generatio

Deep learning has a significantly improved on model performance by learning from large amount of labeled data.However,collecting and labeling tens of thousands of data can be costly.To alleviate the reliance of the recognition system on a large amount of labeled data,zero shot learning has become an important topic in machine learning.Zero shot learning leverages semantic information shared among categories to bridge the gap between visual and semantic features,enabling models to identify novel samples that they have never seen before.Although zero shot learning algorithms have made great progress,they still leave limitations.To address these problems,this paper focuses on zero shot learning based on feature embedding and generation,and the research content is as follows:(1)Embedding-based methods uses a single cross-modal mapping approach to connect the visual and semantic features,which obtains great classification performance.However,visual features and semantic features are different modalities,relying on a single projection function does not adequately fit the complex relationship between them.To solve this problem,this paper proposes bidirectional feature embedding and regression network for zero shot learning.Specifically,this paper designs an embedding and regression framework to improve the embedded features discriminability by returning the learned features from the embedding model to the original feature space in a knowledge migration manner,while alleviating the semantic information loss problem.Secondly,to tighten the semantic relationship between visual and semantic modalities,an embedding regression network including a visual embedding regression network and a semantic embedding regression network with different modalities is designed.The two networks complement each other and collaboratively enhance the discriminative expression of embedding features.Besides,regularization technique and semantic consistency constraint are used to stabilize the model’s learning and improve the features semantic consistency,respectively.Through experimental research,compared to the baseline algorithm,the proposed algorithm improves the comprehensive indicators of the four standard datasets AWA1,AWA2,CUB,and SUN by 2.9%,2.0%,0.6% and 0.6%,respectively,which verifies the effectiveness of the proposed algorithm.(2)Generative-based methods improves the model performance by generating fake visual features of unseen classes,which alleviates the data-imbalance problem between seen and unseen classes during the training phase.However,there is a large amount of semantically irrelevant redundant information in the visual samples,which is likely to interfere with the model learning and thus reduce the discriminative ability of the features.To address this issuse,this paper proposes redundancy-free feature generation network for zero shot learning.Specifically,this algorithm employes a variational autoencoder to generate fake visual samples of unseen classes conditioned on the class-lever semantic information to improve the model’s unseen classes recognition performance.To relieve the redundant information in the visual samples,a feature de-redundancy module is proposed to map the real and fake visual features to a redundancy-free semantic space for eliminating redundant information from visual features.At the same time,semantic regularization constraint is introduced to improve the semantic consistency of non-redundant features.Finally,to mitigate the problem of lost information caused by the process of feature redundancy,a generative adversarial network is introduced to enhance the discriminative property of the learned non-redundant features and further improve the quality of the generated features.The proposed algorithm obtains the best experimental results among the compared algorithms with the comprehensive indicators of 68.7%,68.1%,65.2% and 87.5% on the standard datasets AWA1,AWA2,CUB and FLO,respectively.Moreover,the effectiveness of the proposed algorithm is verified by the feature space visualization study.In summary,bidirectional feature embedding regression network and redundancy-free feature generation network are put forward to relieve the limitations of existing embedding and generation methods.Compared with the existing advanced methods,the proposed algorithms obtain excellent recognition performance in all four major datasets.