Research On Trademark Image Retrieval Application Based On Improved MobileNet

In order to implement the outline of building a powerful country with intellectual property rights proposed by the CPC central committee and the state council,the china national intellectual property administration and intellectual property offices at all levels are vigorously carrying out digital transformation of intellectual property rights.In this process,how to safeguard the legitimate rights and interests of intellectual property rights has gradually become an important research direction of technological innovation.Trademark rights,as an important part of intellectual property,are protected by producers and operators through trademark registration to safeguard their legitimate rights and interests.Among them,intelligent trademark retrieval has become a major challenge for trademark registration due to its high cost and complex process.In the past,people attempted to achieve intelligent trademark retrieval by extracting artificial or content features,but in the experimental process,it was found that there were still problems such as slow retrieval speed,low accuracy,and unstable results.In response to the above issues,this article proposes a trademark image retrieval algorithm based on convolutional neural networks.The specific research content is as follows:(1)In order to solve the problem of unstable image feature information extraction caused by small target graphics and text compound trademarks,this paper proposes a two-stage trademark retrieval model.The specific implementation method is to use the improved YOLOv3 model to detect and locate effective graphics in the trademark image,and extract feature information from the corresponding image area through a feature extractor.Finally,through experimental data analysis,the two-stage trademark retrieval model proposed in this article can significantly improve the accuracy of trademark retrieval.(2)In the current database with a large number of registered trademark images,using deep convolutional network models for feature extraction of trademark images may lead to rapid consumption of computing resources and a significant increase in model size.This article adopts the lightweight convolutional neural network model Mobile Net as the baseline model,and then optimizes the Mobile Net model from three aspects: model accuracy,model parameter size,and model resource consumption.While improving the accuracy of the Mobile Net model,it further compresses the model volume and speeds up model operation.The experimental results show that the optimized Mobile Net model has corresponding improvements in accuracy,computational speed,and model size.Finally,this paper uses milvus vectorization database as the image feature storage database,uses a variety of vector distance calculation formulas,calculates the distance between the feature vector to be retrieved and the feature vector stored in the database,and quickly calculates the distance between the vectors to get the image retrieval results.