Research And Design Of Trademark Retrieval System Based On Deep Learning

Trademarks are an important way to distinguish a company' product or services from others and are an important resource for companies.In recent years,it has become challenging to retrieve desired trademarks from a large number of trademarks due to the increasing number of trademarks.The existing methods usually retrieve the similar trademarks by extracting and the features of query trademarks directly,which ignores the problem of mutual interference between various categories of trademark features and the problem of poor retrieval results for low-resolution query trademarks,seriously affecting the user experience of the trademark search system.Aiming at the above-mentioned problems in trademark retrieval,this paper researched the trademark retrieval method from the aspects of hierarchical search strategy and image super-resolution reconstruction,and designed a trademark retrieval system based on deep learning.The main work and innovations of the paper are as follows.1.Researched the effectiveness of hierarchical strategies for trademark retrieval.Interference among various categories of features in existing trademark retrieval methods is more obvious,and existing retrieval algorithms are not effective in reducing this interference.In order to address this issue,we proposed a trademark retrieval method using a hierarchical search based on deep learning.The method includes two stages: coarse-classification and fine-retrieval.The candidate category set and deep features of the query image are obtained through coarse-classification.Fine-level retrieval mainly filters out better retrieval ranges from the image feature library by the candidate category set,and performs similarity matching and ranking according to the retrieval ranges.Our method not only weakened mutual interference between features of different categories but also discussed the effect of candidate category ranges on retrieval accuracy.2.Researched the effectiveness of image super-resolution reconstruction techniques for trademark retrieval.To address the negative impact of real-life,low-resolution query trademarks and datasets on retrieval accuracy,we proposed a trademark retrieval method based on image super-resolution reconstruction,which adds the image super-resolution reconstruction technique in the trademark retrieval system as image preprocessing of the trademark.The method can reconstruct the low-resolution trademark to high-resolution,thus improving the trademark retrieval accuracy.3.Designed a trademark retrieval system based on deep learning.We designed and optimized the trademark retrieval system from the functional and performance requirements,and implemented and analyzed the system algorithm module.