Zero Sample Image Classification Based On Semantic Knowledge

Zero-shot image classification establishes the bond between the seen category and the unseen category through semantic knowledge,so that it can identify the categories that have not been seen in the training stage.It has become one of the main research problems in the computer vision.The established bond includes semantic attributes and semantic word vectors.Semantic attributes are semantic descriptions of images.Semantic word vectors are text features corresponding to object category.The current zero-shot image classification algorithm studied attribute learning and mapping function learning from visual to semantic.However,the existing method ignored the similarity metric calculation and the construction of semantic knowledge.This paper has improved its deficiencies.The main work includes:(1)Aiming at the problem of different attributes having the same weight for classification decisions,the zero-shot image classification algorithm based on entropy attribute weight learning is proposed.The joint embedding model of vision and semantics is used by the fully connected layer.Then,the entropy method is introduced to learn the attribute weight,so that the attribute weighted distance function is obtained and used for the nearest neighbor search.The experimental results show that the improved method enhanced the accuracy of the model and proved that selecting of the different embedding spaces will also affect the recognition accuracy.(2)Aiming at the limitation of single auxiliary knowledge descripting target,the zero-shot image classification algorithm based on adaptive weighted fusion feature is proposed.The neural network is used to fuse text features and semantic attributes.Then,the particle swarm optimization algorithm is introduced to determine the weight of feature fusion,which can enhance the richness of semantic information.The weighted fusion features are used for zero-shot image classification.The experimental results show that the improved method improved the recognition accuracy,robustness of the classification model and enhances the robustness of semantic knowledge.(3)Aiming at the problem of the fixed metric function in similarity calculation.The improved semantic autoencoder classification algorithm based on metric learning is proposed for the shortcoming.Semantic autoencoder is used to construct the joint embedding model of vision and semantics.Then,the metric function of adjustable parameters is obtained by learning the original data.And the obtained metric function is used to calculate the similarity between features.The experimental results show that the metric learning improved the distribution of the features and made the feature distribution inside class more compact,reducing the error rate of the classification model.Aiming at the insufficient of metric methods and semantic knowledge,this paper proposed the improved algorithms based on entropy method,particle swarm optimization algorithm and metric learning,which improved the accuracy and robustness of the model and enhance the richness of semantic information.