Research On Zero-shot Learning Algorithms Based On A Learnable Deep Metric

Zero-Shot Learning(ZSL)is proposed to solve the problem that classifiers cannot accurately recognize categories that lack labels.The purpose of ZSL is to allow deep learning models to accurately recognize never-before-seen categories in the absence of training samples.Most of the existing ZSL models map semantic features into the visual space,then the predefined fixed measures(e.g.,Euclidean distance)shall be used to obtain the relationship between the visual and semantic features to complete the matching of unseen class labels.However,the use of predefined fixed measures yields the difference between the two,a relationship that is not learned and therefore has limitations.In the cross-modal mapping process,there are differences between the manifolds of semantic features and visual features,which may lead to the Semantic Gap problem.Existing ZSL methods based on spatial embedding mostly use the global features of an image as the embedding of visual features,neglecting the important role of discriminative regional features.A Zero-shot Image Classification based on a Learnable Deep Metric(ZIC-LDM)algorithm is proposed to address the limitations of predefined fixed metrics in ZSL and the semantic divide between visual and semantic features in the cross-modal mapping process.The ZIC-LDM algorithm uses end-to-end learnable depth metrics instead of predefined fixed metrics and uses common space embedding to mitigate the semantic gap issues.Considering the important role of discriminative regional features,a ZSL algorithm based on feature fusion,namely Feature Fusion with Discriminative Region(FFDR),is proposed.The FFDR algorithm fuses discriminative regional features with global features and uses learnable depth metrics to learn the relationship between the fused features and semantic features.Compared with baseline,ZIC-LDM was increased by 1.0% and 5.5% in the accuracy of Aw A1 and Aw A2,meanwhile,in the harmonic mean of Aw A1,Aw A2 and CUB respectively by 4.3%,12.6% and 4.6%.Respectively,FFDR was increased by 2.1%and 1.1% in the accuracy of Aw A2 and CUB.Figure 22;Table 6;Reference 71...