Trademark Image Retrieval Based On Deep Learning

With the rapid development of the economy,the number of enterprises has increased substantially,and the number of enterprise trademark registrations has been increasing,which makes it more difficult for the trademark management department to manage the trademark of the enterprise.However,the management of trademark image management by trademark management department still adopts the method of "classification number",which is not only the low accuracy of retrieval,but also the retrieval efficiency is not high.Traditional trademark image retrieval technique is often used to extract the global characteristics of trademark image,and then use these features to match the image recognition.But because the logo image often has a strong abstractness and complexity,making use of the global characters of the manual design trademark image retrieval will bring"semantic gap"problem,the retrieval precision of image retrieval system is not high,the recall rate is low.In recent years,with the optimization of deep neural network model,Hinton et al.brought new solutions to the problem of image retrieval.This paper tries to use in the field of deep learning techniques to solve the problem of trademark image retrieval,and through the experiment to determine whether a deep learning technology can sol've the problem of "semantic gap" in the trademark image retrieval.The main work and contributions of this paper are as follows.First of all,by collecting a large number of trademark images,a large image library of trademark is set up,and the image library is divided into six feature classes by using the geometric features of the trademark image.Its main purpose is to use the feature extraction of trademark image,and also helps researchers to study the trademark image later.Secondly,the research in the field of deep learning technology,and puts forward a deep convolution based trademark image feature extraction method of neural network,and through it with the Bag of the Features in the model by using SIFT image characteristics of artificial design method,the final results show that trademark feature extracting convolution neural network method has a higher retrieval accuracy.Finally,this paper will expand the query technology for image retrieval system,and propose an improved average query expansion(AQE)method.