Few-shot Logo Detection Research

As an important part of corporate identity,logo has multiple functions such as brand identification,brand image enhancement,product promotion and enterprise competitiveness enhancement,which is of great significance to the development and success of enterprises.The purpose of Logo detection is to detect the Logo in the image,determine the category of the Logo,and mark its location.As logo detection has important applications in brand protection,advertising monitoring,commodity identification,social media analysis,intelligent transportation and other aspects,more and more logo detection methods have been proposed,and these logo detection methods all require a large amount of data to train the model.However,due to the low frequency of logos of some emerging brands and the high cost of manual data annotation,the amount of data used for training is limited,which promotes the development of few-shot detection.Besides,due to the multi-scale logos in a single image,and the boundary of some logos is not clear compared with general object detection,which further increases the difficulty of few-shot logo detection.Therefore,this thesis will carry out research on few-shot logo detection.The contribution of this work can be summarized in the following two aspects:(1)In order to solve the problem that the accuracy of base class decreases significantly in the detection process,this thesis proposes a Double Classification Head based Few-shot Logo Detection method.Unlike traditional few-shot object detection,logo objects are multi-scale,and some categories of logo objects are small or integrated with the background,these factors increase the difficulty of few-shot logo detection.In this paper,a Double Classification Head based Fewshot Logo Detection method is proposed to solve the above problems.Specifically,this thesis uses the balanced feature pyramid and deformable Ro I pooling to adapt to multi-scale logo objects and improve the feature extraction ability.In addition,this thesis also designs a double classification head for few-shot logo detection,which is used to reduce the impact of novel classes on base classes during classification.The experiments are designed on four logo datasets,under the 1-shot of Flickr Logos-32,this method is better than baseline with up to 8.3% n AP on novel classes and15.5% b AP on base classes.The results illustrate the effectiveness of this method.(2)The few-shot logo detection method is difficult to detect the multi-scale logo objects,therefore,this thesis designs a Two Branch based Few-shot Logo Detection method.Due to the limit training sample of the few-shot detection method and the single object scale for training,it is difficult for the model to extract multi-scale features.In addition,the multi-scale characteristic of the logo object increases the difficulty of few-shot detection.In order to solve this problem,this thesis proposes a few-shot logo detection method based on two branches.The resized objects input refinement branches for training.The trained RPN and detector head are used to fine-tune the RPN and detector head of the Fast R-CNN.In addition,the balanced feature pyramid is added to the backbone to further enhance multi-scale feature extraction.The experiments are designed on Flickr Logos-32 and Food Logo Det-1500-100 datasets,for 1-shot,the accuracy of novel classes about these two datasets are 43.0% and 31.3%,respectively.