Research On The Method Of Zero-Shot Learning Based On Embedding Space

Traditional image classification tasks require a large amount of labeled data for training,but in real life,data collection and labeling are very difficult.Therefore,the zero-shot learning of how to recognize objects when there are insufficient samples or even no samples has become a research hotspot.Zero-shot learning is a branch of transfer learning,which aims to classify categories that have not appeared in the training process.The current mainstream research directions include zero-shot learning based on semantic embedding space and Zero-Shot Learning based on visual embedding space.This thesis researches and improves the algorithms in these two directions respectively.The main contents of the work are as follows:(1)Semantic Auto Encoder(SAE)is a classic algorithm based on semantic embedding space,which can obtain good classification results through mathematical analytical solutions.However,the SAE algorithm has the too strong constraint on the objective function,which makes the model prone to overfitting.In response to this problem,this thesis chooses to perform Linear Optimization on the SAE,adding a linear discriminant to its objective function to broaden the constraint of the model.In addition,a reasonable and flexible measurement method can accurately measure the relationship between features,thereby improving the classification performance of the model,while the traditional measurement method adopted by the SAE is artificially preset,so it has certain limitations and unicity.For this,this thesis introduces a neural network-based measurement module,which is based on a data-driven method for learning,without manual selection of a fixed measurement method,so the learned measurement module is more flexible and robust.After the Semantic Auto Encoder based on Linear Optimization and Metric Learning proposed in this thesis is compared and tested on the Aw A and CUB data sets,it is found that the proposed algorithm has better classification effect than the SAE on both data sets.(2)In view of the common domain shift problem in the mapping model of the semantic feature embedded in the visual space,this thesis chooses to add the Reconstruction Constraint in the process of semantic feature embedding,so that the embedded features can retain more effective information and alleviate the domain shift.In addition,for the lack of the expressive ability of the single attribute feature or the word vector,this thesis chooses to integrate the word vector with the attribute feature to enhance the expressive ability of semantic features.After the Deep Embedding based on Semantic Fusion and Multi-Reconstruction Constraint for Zero-Shot Learning has comparative experiments with Aw A and CUB data sets,it is found that the proposed algorithm has achieved good classification results on both data sets.