Zero-Shot Fine-Grained Object Classification Based On Latent Attribute Dictionary Learning

Zero-shot fine-grained object classification aims to distinguish subordinate categories that even unseen through some other seen categories within an entry-level category,which obtains an important position in the fields of new retail,face identification and intelligent management system.Zero-shot learning(ZSL)recognizes an object instance from a new category never seen before with the help of semantic cues that transfer knowledge between seen categories and unseen categories.Typical ZSL methods require that only one object appears in a test image.Human-defined attributes are mostly utilized as the semantic cues for direct knowledge transfer.Such settings lead to the insufficient exploitation of attributes,resulting in inadequate feature descriptions.The human-defined attributes are not exhaustive and discriminative enough.The typical setting that each test image contains only one object hinders the real-world applications of fine-grained recognition.To this end,we focus and do research on the followings in this paper:Firstly,we do research on theories related to zero-shot fine-grained object classification that based on attributes learning.The first stream utilizes human-defined attributes for knowledge transfer so that the untrained categories can be recognized correctly.However,human-defined attribute suffers problem of inadequate descriptive.The second stream explore latent attributes for zero-shot classification.These methods study latent attributes through jointly learning with human-defined attributes and obj ect labels,which makes the latent attributes both discriminative and semantic-preserving.Latent attributes can accommodate with human-defined attributes to alleviate the problem of inadequate feature description and they are more descriptive.Typical datasets of zero-shot classification are discussed.We perform experiments on these datasets to validate the effectiveness.Secondly,we explore and propose a similarity-specific latent attribute learning model for typical zero-shot fine-grained object classification.To solve the extremely challenging problem of low inter-class invariance and high intra-class difference,we propose to learn for each visually indistinguishable feature a similarity-specific dictionary with discriminative latent attributes.In general,the instances are modelled through Gaussian Mixture algorithm to be departed into clusters.We learn for each cluster a dictionary through jointly training with other two spaces,which leads to similarity-specific latent attributes for specific kind of feature.The classification is done by the combination of all the learned dictionaries.The distinctive training phase benefits the performance of zero-shot fine-grained obj ect classification.Thirdly,in the scenario of recognizing multiple fine-grained objects within a test imagery,we propose a novel framework in this paper.To handle the signal decimation problem of coarse-grained object localization and recognition in the process of remote sensing imagery processing,we propose a hierarchical DeepLab v3 that fuses the low-level image features and high-level image features to alleviate signal decimation of small objects.Generalized zero-shot algorithm is adopted for object classification to address the limitation of typical zero-shot classification methods that require test instances can only come from unseen categories,which is more adaptable in the scenario.Extensive experiments are done to validate the effectiveness.