A Real-time Trademark Detection Method Based On Improved YOLOv3

With the advent of the knowledge economy,China has paid more and more attention to the protection and application of intellectual property rights.As an indispensable part of intellectual property rights,the importance of trademark protection is obvious.Searching for trademarks is a key technology for trademark protection and application.Mobile Internet technology is developing rapidly,which has penetrated into all aspects of people’s lives.And the snap and search has become a very urgent need.In the snap and search of trademark,real-time detection is a key step.With the increasing number of trademarks,trademark infringements in various industries have emerged endlessly.Searching for similar trademarks in such a huge database,real-time detection technology plays an important role.Trademark detection technology is essentially the application of object detection technology in the field of trademarks.YOLOv3 is a widely used real-time object detection model and has been successfully applied in some fields such as license plate detection.However,if YOLOv3 is directly used for trademark detection,there are problems of complex background interference and large differences in trademark size,which seriously affects the effectiveness of trademark detection.According to the characteristics of the trademark,we have conducted in-depth research on trademark real-time detection methods based on YOLOv3.The main research contents include:1)The related methods and technologies of real-time trademark detection are studied.We classified and compared real-time object detection methods based on deep learning,and selected YOLOv3 as the object of improvement.Aiming at the problems of trademark detection in natural scenes,the theory of feature pyramid network and attention mechanism is studied.2)The multi-scale detection network based on improved feature pyramid is studied.The analysis of trademark data shows that the size of trademarks varies greatly,and most of them are small trademarks.Because YOLOv3 is not robust enough to the trademark size,there will be a phenomenon of false detection.Therefore,a new detection layer containing a large-scale feature map is added to form a new feature pyramid work with a larger scale range,which forms a better multi-scale detection network in this paper.This not only strengthens the detection effect of small trademarks,but also improves the robustness of the detection model to trademark size.3)A real-time trademark detection method based on improved YOLOv3 is proposed.On the basis of improved feature pyramid network,this method introduces channels and spatial attention mechanisms in the feature extraction network in order to solve the problem of complex background interference to trademark detection,so that the network pays more attention to the features of foreground trademarks.In addition,residual connections based on second-order terms are used to enhance the non-linearity and generalization performance of the feature extraction network.4)The effectiveness of the proposed method is verified.This paper uses the Flickrlogos-32 trademark dataset to verify the proposed method LOGO-YOLOv3.Experiments show that,compared with the original YOLOv3,LOGO-YOLOv3 achieves higher detection accuracy while maintaining real-time capability.