Research On Zero-shot Recognition Based On Deep Learning And Attention Mechanism

Traditional deep learning methods usually rely on a large number of manually annotated data in the learning and training process,and the data used for testing are often separated from the existing categories of training data.However,in practical application,it is difficult to obtain a large amount of annotated data,and it is more consistent with human cognitive habits to use the information of existing categories to identify unknown categories.Inspired by the process of human cognition,zero-shot learning can use the information of seen class and additional auxiliary information to complete the transfer from seen class to unseen class.In this paper,two zero-sample image classification methods based on deep learning are proposed to solve the problems of attribute relation learning,visual feature learning and domain migration.1.Aiming at the problem that the traditional zero-shot learning method based on spatial embedding only uses pre-trained network to extract fixed image features and only carries out embedding in a single space,we proposed a zero-shot learning method which can automatically learn multi-level visual features and jointly train semantic embedding and latent feature embedding.Based on the pre-trained image feature extraction network,this method adaptively generates attention about channel by fusing image features of different levels,and jointly trains the embedding from visual space to semantic space and the embedding from visual space to latent feature space.At the same time,a discriminative region extraction network is introduced,which can independently find the discriminant subject in the image and perform appropriate cutting,so that the network can get more accurate image features.In the zero-shot classification stage,the prototype is established in semantic space and latent feature space respectively,and the zero-shot image classification is completed by searching and testing the category prototype closest to the sample.2.A zero-shot learning method based on adaptive prototype construction is proposed to solve the problem of domain offset in traditional spatial embedding methods.In this method,channel attention module and spatial attention module are introduced in image feature extraction to automatically learn feature weights,encouraging positive features and suppressing unimportant features.An end-to-end network is constructed to jointly train the image feature extraction network and the embedding network.During the training process,the parameters of the pre-trained image feature extraction network are constantly fine-tuned to obtain the image features that retain the correct semantic information and eliminate the interference of unimportant information.In the zero-shot classification stage,a method of directly adaptive prototype construction in hidden attribute space is proposed by using known seen class prototype in semantic space,which solves the problem caused by domain shift well.A large number of experiments on three datasets Aw A2,CUB and SUN show that the two methods proposed in this paper achieve the most advanced performance,effectively alleviate the problem of domain shift in zero-shot learning,and improve the accuracy of zero-shot image classification.