Research On A Trademark Image Retrieval Method Based On Deep Supervised Hashing

With the accelerating process of economic globalization,enterprise legal persons and producers have gradually increased their brand awareness and appeal for trademark right confirmation.Trademark examination is an important link of trademark right confirmation and protection.At the same time,trademark retrieval is the key means of trademark examination and application.In order to meet the rapidly growing demand for trademark retrieval,both the efficiency and accuracy of trademark retrieval need to be further improved.How to construct an efficient and accurate trademark retrieval method has become a key problem to be solved urgently.To solve this problem,scholars at home and abroad proposed several trademark retrieval methods based on deep hashing,which mapped high-dimensional trademark features to Hamming space for distance calculation,in order to solve the problem of poor trademark feature matching results,and improve the retrieval accuracy and efficiency of trademarks.These methods make use of the supervised hashing method which maintains the pairwise similarity,and achieve good results in the efficiency of trademark retrieval,but they can not solve the problem of the imbalance of trademark data.Aiming at the problem of unbalanced trademark data,this paper proposes an improved trademark image retrieval method.The main research work is as follows:(1)The relevant technology of trademark retrieval is studied.The background and significance of trademark retrieval method are analyzed.The existing domestic and foreign trademark retrieval systems are investigated and analyzed.The classical trademark retrieval method based on the bottom feature and the trademark retrieval method based on the depth feature are studied respectively,and the experimental results show that the retrieval efficiency and retrieval precision of the classical trademark retrieval method need to be further improved.(2)An improved trademark image retrieval method is proposed.At present,the supervised hash method has achieved good results in the efficiency of trademark retrieval,but it is faced with the problem of unbalanced trademark data.The proposed method firstly uses the properties of Hadamar matrix or Bernoulli distribution to generate well-separated trademark hash centers,then associates trademark labels with them so that they have corresponding semantic centers,and finally optimizes the central similarity between trademarks through trademark center quantization.This method can effectively reduce the dimension of trademark features and improve the retrieval accuracy.(3)Experimental demonstration was carried out.In public data sets and manual classification of a trademark data set on the contrast experiment,the experimental results show that the proposed trademark retrieval method based on the depth of the hash compared to classic trademark retrieval method,on the retrieval precision of trademark and the retrieval efficiency has certain of ascension,and the depth of the existing supervision hash algorithm is more suitable for unbalanced logo image,better robustness.