Research On Trademark Retrieval Methods Based On Aggregation Similarity Capsule Hashing

With the booming development of 5G,big data technology,industrial internet and other new generation information technology,the number of global trademark data is expected to reach 65 million by 2035.The complex structure,variety and ZB magnitude of image data attached to the drastic increase in the number of trademarks will undoubtedly demand more severe accuracy and efficiency for trademark audit and search.It is increasingly important to quickly and accurately search for similar and confusing trademark images in the massive trademark database,making trademark search one of the essential technologies for protecting users’ rights and ensuring trademark exclusivity at present.One of the most advanced hashing methods currently available is the use of deep neural networks,particularly convolutional neural networks(CNN),to obtain image hash codes for fast image retrieval.However,CNNs cannot handle ambiguity well,and a large amount of information is lost in the pooling layer leading to the inability of CNNs to learn the spatial hierarchy of trademark images;in addition,trademark retrieval methods based on convolutional neural networks have difficulties in handling massive and high-dimensional nearest neighbour feature map searches.At the same time,the existing pair/triplet trademark hashing retrieval methods cannot reconcile the contradiction between data imbalance and accuracy enhancement,and suffer from the deficiency of not being able to capture the global semantic relationships of trademarks.In this paper,based on the characteristics of trademarks themselves and comparing the limitations of existing trademark image retrieval methods in the design of the underlying algorithm for processing trademark images,by improving on the basis of traditional image hash retrieval,making in-depth improvements to its classical algorithm and combining the advantages of capsule network technology in extracting deep semantic features,an image retrieval algorithm based on a deep hash learning framework capable of quickly processing images that are sensitive to relative spatial structure and have imbalance is designed,with the main work and innovations as follows.1)In view of the powerful capability of capsule network algorithm in deep semantic representation learning,and in response to the traditional convolutional neural network’s inability to meet the requirements of fine-grained trademark retrieval due to the lack of considering the extraction of the relative spatial location of trademark element parts,this paper proposes a trademark retrieval method based on the extended capsule network as a framework,and the underlying algorithmic mechanism for the extraction of the spatial hierarchical information of trademark element parts is designed to capture the key feature information of the spatial hierarchical relationship of trademark element parts,in order to further enhance the network’s ability to represent trademark images and improve the accuracy of trademark retrieval tasks.The aim is to capture the key feature information of the spatial hierarchical relationship of the trademark element components,in order to further enhance the network’s ability to characterize trademark images and improve the accuracy of the trademark retrieval task.In comparison with other classical deep convolutional networks and primitive capsule networks,this method can further improve the accuracy of the trademark retrieval task by extracting the complementary features of the relative spatial relationships of the elemental components of the trademark image,compared with four datasets of MNIST-10,CIFAR-10,Flicker Logos-32 and Drink Logos-50.2)Based on the analysis at the level of index matching architecture of trademark image retrieval,combined with the natural advantages of low storage and high efficiency of hash retrieval methods,to address the traditional hash methods in the process of improving retrieval accuracy,the lack of consideration of data imbalance characteristics,therefore,this paper introduces the concept of hash aggregation cluster core,and proposes a trademark retrieval method based on aggregated similarity depth hash,by pre-determining the trademark aggregation cluster core The aim is to enable the retrieval model to capture the global semantic relationships of trademark images.At the same time,it improves the network’s ability to represent the data differently without changing the data balance,prompting the aggregation of trademarks within classes and keeping them away from each other.In comparison with other advanced hash trademark retrieval methods,the aggregated similarity depth hash network model proposed in this paper,by capturing the global semantic relationships of trademarks and effectively processing unbalanced trademarks,solves the data imbalance and retrieval By capturing the global semantic relationships of trademarks and effectively processing imbalanced trademarks,this paper solves the contradiction between data imbalance and retrieval accuracy from the underlying logic,thus improving the accuracy and retrieval response efficiency of trademark retrieval models.3)Given that the capsule network has better image representation extraction ability and sensitive characteristics of relative spatial location of element parts than the traditional convolutional neural network,and the excellent performance of global representation and high retrieval efficiency of the aggregated similarity depth hash network,the two are organically combined to form a new type of multi-fast and cost-effective trademark retrieval network.Therefore,based on this concept,this paper proposes a trademark retrieval method based on aggregated similarity capsule hashing,and introduces a sparse coding mechanism to reduce the redundancy of coding between hash bits and lighten the model in view of the shortcomings of this network model with many encapsulated components and large parameters;at the same time,a new fusion loss function is designed for the cascaded fusion model,which enables the cascaded fusion model to achieve dynamic adjustment of its parameters by modules update,thus improving the convergence speed of the overall model.Through multiple comparisons on the CIFAR-10,Flicker Logos-32 and Drink Logos-50 datasets,this paper’s method is more scientific and feasible compared with other classical hash retrieval methods.