Research On Zero-shot-learning Based On Fusion Of Semantic Information

In the field of image recognition of artificial intelligence,recognition and classification is still the key to various application tasks.It is the basic requirement on all kinds of application scenarios to determine the category of objects accurately.In the recent years,with the development of deep learning,the method of deep learning image recognition can achieve higher accuracy and faster speed than human beings in various fields,such as face recognition,medical detection and industrial defect recognition.However,in order to achieve preset recognition effect,the model needs to have excellent fitting ability.It not only requires large-scale annotated data samples but also needs to ensure that the sample categories of training set and test set are consistent.However,in practical applications,it is difficult to obtain some kinds of image sample data.On the other hand,the new samples of unknown categories make the model need to be retrained in order to recognize the new categories.Therefore,the use of prior knowledge realizes zero-shot-learning of identifying unknown categories through known category training,which has practical research significance.At present,zero-shot-learning method is mainly based on attribute features or semantic features as the category prior features to build the mapping between the image features and the prior category features,which completes the domain migration from the known category to the unknown category.However,in the general zero-shot image classification model,the image feature extraction module lacks a screening mechanism for key and non-key visual features used to match the prior category attributes.In addition,using only attribute features or semantic features makes description information insufficient.As a result,it is difficult to match image features and semantic features,which affects the accuracy of the zero-shot-learning model.To this end,this paper proposes a systematic zero-shot-learning method.Based on the original mapping model of image features and prior category features,two main improvements are proposed.One is the image feature extraction network based on the attention mechanism,which is transformed on the basis of the convolutional neural networks including VGG-19 and Res Net-34.The combination of the above model features is carried out by means of spatial attention focusing.Its effect needs to be verified.The second is a semantic information fusion method based on matrix decomposition.Three types of prior semantic knowledge word2 vec,Glo Ve and fast Text are acquired through targeted training.Through matrix decomposition of the attribute features,the attribute decomposition vector and the semantic feature vector are matched in dimension.This achieves vector fusion,and then can carry out information exchange.Its effect needs to be verified.In the zero-shot image classification task,the recognition accuracy of the unknown category is the standard.Experimental results show that comparing various image feature extraction networks,the image feature extraction model Res Net-A based on Res Net-34 and attention mechanism has the best effect;comparing the fusion of various semantic features and attribute features,the fusion semantic feature information which mixed by Glo Ve semantic features and attribute features is the best.