Local-global Joint Representative Hashing For Fine-Grained Image Retrieval

Hashing based fine-grained image retrieval maps the fine-grained dataset to the Hamming space to generate binary hash codes.Fine-grained datasets refer to datasets in which all images belong to the same meta-category,such as commodity dataset,bird dataset,and vehicle dataset.Since all images come from the same meta-category,images in different categories are very similar.And due to different poses and view-points,fine-grained images belonging to the same category will also have large intra-class differences.In order to identify fine-grained images,it is necessary to calculate the local features of the fine-grained images,and the calculation of local features involves a relatively large time complexity.Thus,how to effectively retrieve the large-scale fine-grained images becomes extremely challenging.Due to the efficient storage and calculation of hashing,it has received more and more attention in retrieval task.Hashing maps images to binary hash codes and uses Hamming distances to calculate the distances between samples.In order to efficiently retrieve fine-grained images,this paper studies fine-grained image retrieval based on hash technology.After summarizing the current status of fine-grained hashing methods at home and abroad,this article combines the local-global joint representation,central loss function,hard sample mining,and outlier processing to solve the fine-grained image hashing.The main contributions are as follows:(1)A sub-region localized hashing for fine-grained image retrieval is proposed.This algorithm proposes a local-region localized module,which locates key points in evenly divided sub-regions,and uses the key points as anchor points to calculate local features.Since the local features are located in each uniformly divided sub-region,the local features are dispersive and diversified.In addition,the algorithm uses the orthogonalized central loss function to optimize the hash network.Through the central loss function,the learned hash codes are compact in each class.And through the orthogonalization for hash centers,the learned hash codes are dispersive cross different classes.This paper uses four widely-used fine-grained datasets to conduct experiments,which proves the effectiveness of the proposed method.(2)A fine-grained hash algorithm based on local-global feature fusion method is proposed.This method performs hash coding on fine-grained images from the perspective of local-global feature fusion.The local features are calculated by the channel-spatial attention mechanism,and the global features are obtained by global average pooling of the feature map.Then,the local features and global features are concatenated to obtain the local-global feature fusion representation.The hash codes are learned from the fused features.To optimize the hash network,a pair-weighted based loss function is proposed to mine hard samples,which makes the learned hash codes more discriminative.However,outliers inevitably exist in the learned hash codes.This method proposes a self-adaptive module,which can detect the outliers in the hash codes and correct the outliers.A large number of experiments can prove the effectiveness of the proposed method on two fine-grained datasets.