Logo Detection Neural Network Based On YOLOv5

In an enterprise's growth process,brand logo serves as its important symbol.The emergence of increasing number of counterfeit brands has caused extremely serious impact on corporates' images.Thus,brand logo protection serves as one of the core strategies of enterprise growth.To crack down on counterfeit goods and deal with trademark infringement,thus resolving brand crisis,brand equity protection is an important measure for brand development.With the development of deep learning technology and increase in computing power,the usage of deep-learning-based techniques for brand logo detection and localization in images has become a classic issue in the field of computer vision.To be applied practically,logo detection generally has strict requirements for speed and accuracy.Although complex target detection networks can meet the accuracy requirements,they are often unable to meet downstream application requirements in processing speed.In addition,the lack of logo sample data can significantly limit the quality of logo detection.Therefore,training quick and accurate logo detection models with the constraint of limited sample data is a high-priority research topic in this field.In order to solve the problems mentioned above,this paper used YOLOV5(a world's first)for logo detection and carried out 32 types of logo detection experiments on the public Flickr Logos-32 data set.At the same time,Flickr27 data sets(including 27 kinds of Logo,Logo category is not completely consistent with Flickr Logos-32 of the type of Logo)are used to verify YOLOv5's logo detection performance,and compare it with that of the one-phase YOLOv3 spp model and double-phase Faster-RCNN model.Performance metrics evaluated include m AP,processing time,model size etc.On this basis,this paper compares the difference between models as shown by the experimental results.The conclusion is that YOLOv5 is superior to existing models in logo detection due to its 5advantages in Fast detection speed,high accuracy,small model size,easy deployment and compatibility with limited,small datasets.