Abstract:

The existing few-shot image classification algorithms based on metric learning have a low precision of image categorization and weak generalization performance.A few-shot image classification algorithm by dynamic subspace is proposed in this paper.First,a residual neural network is used to extract few-shot image features.The dynamically orthogonalized projection subspaces representing image categories are generated with decomposed image features of various categorizations so as to enhance the difference of features among categories in orthogonalized projection subspaces.Second,a dynamic subspace classifier based on few-shot learning is constructed by fusing the subspace loss function and the cross-entropy loss function so as to enhance the similarity of samples in the same category.The inter-class distance of the subspace is dynamically updated with the change of sampling amount and sample similarity.Finally,the feature vector of the target image is input into the dynamic subspace classifier,and the Euclidean distance square and the soft max function are used to calculate the category probability of the target feature and predict its category.Performance testing is performed on the few-shot data sets such as mini-ImageNet,CIFAR-100 and Pascal VOC2007.The proposed algorithm is superior to the current mainstream few-shot image classification algorithm,and the average classification precision of the proposed algorithm is 2.3% higher than that of the current DSN with good performance under 5-way 5-shot.Experiments show that the proposed algorithm has a strong generalization performance and an anti-interference ability.

Key words: dynamic subspace, few-shot data, orthogonal projection, image classification, residual network