Research On Logo Detection And Recognition Based On Deep Learning

With the continuous development and popularization of computer vision technology in recent years,object detection and recognition technology has demonstrated its convenience and advantages in more and more aspects such as medical and health care,national education,urban transportation,and people entertainment.Logo detection and recognition is an important application of object detection and recognition technology,and it also has application prospects that can not be ignored in the fields of urban intelligent transportation,document retrieval and classification,and commercial advertisement analysis.Most of the traditional logo detection and recognition technologies are based on single logo in the document,that detection and recognition environment is relatively simple and the application research is less difficult.This thesis takes the advanced object detection and recognition methods as the research foundation,takes the detection and recognition of sports brand logo in social networks as the entry point,and researched logo detection and recognition based on deep learning.The work mainly includes three aspects: data set construction,dataset expansion,and deep learning model selection and improvement.Firstly,this thesis researched and analyzed the existing public datasets in the field of logo detection and recognition,and found that the existing public datasets couldn't satisfy our specific research and application requirements.Therefore,we adopt the web crawler and manual annotations method to construct FlickrSportLogoss-10 dataset,which contains ten sports brand logos suitable for our research.Secondly,the training of deep learning model requires a lot of labeling data,but marking the data set is very tedious and will consume a lot of manpower and material resources.Therefore,this thesis used the image synthesis technology to expand the FlickrSport Logos-10 data set with 100000 images,and we proved the effectiveness of this data augmentation method in later experiments.Finally,based on FlickrSportLogos-10 dataset,this thesis selects three kinds of algorithms for research and comparison,and carries out relevant experiments: running result,running time,model visualization and so on.Finally,it summarizes the different application scenarios for the three logo detection and recognition algorithms.In summary,the research in thesis fully proved the feasibility and superiority of the logo detection and recognition based on deep learning,and verified the effectiveness of the image synthesis technology for logo data augmentation.Compared with the traditional research,logo detection and recognition technology studied in this thesis has more complicated detection and recognition background and is more difficult to detect and recognize,and the method we used is more effective.