Research On Logo Detection Algorithm Based On Deep Learning

The brand logo represents the image of the company,which not only has commercial value,but also contains the cultural value of the brand.Besides,the logo is a symbol of the company,which embodies a vital role during the development of the company.With the social becoming increasingly information-oriented,more and more trademarks are being imitated by counterfeit brands,which seriously endangers the brand image and affects the revenue of the brand.In order to protect brand trademarks to maintain corporate image,it is urgent to combat counterfeit goods and deal with infringing trademarks.In the implementation of the application and protection of intellectual property rights,the accurate finding of brand logo is one of the key techniques for the protection and application of trademarks.In recent years,with the development of image processing technology and the improvement of computing power,image processing technology has been widely applied to object recognition problems.There has been widespread interest in how to use object detection algorithms to identify similar trademarks in trademark infringement incidents of brands.Commodity image retrieval technology can detect commodity information in real time through commodity images and list products with similar logos.In contrast,merchandise Logos in natural scene are characterized by variable shapes,too small targets and complex image backgrounds compared to ordinary detection targets,all of which seriously affect the recognition effect of logo detection.Based on the above background,we takes the logo target in the actual scene as the research object,in order to improve the accuracy of target recognition and research on the logo recognition method based on deep.The main work of this paper are as follows:1)For the situation that the target size in the logo data varies greatly and there are many small targets,this paper proposes a detection algorithm that adapts to multi-scale training.A Cascade R-CNN network framework with a cascade structure is selected as an improved model,and a feature pyramid FPN structure is introduced in the RPN sub-network,so that the model can fuse multi-layer features for training.The Swin Transformer is used instead of the traditional CNN as the feature extraction network,and the self-attention mechanism is used to extract more feature information of the small logo target in the original image to improve the algorithm performance in recognizing multi-scale objects and increase the robustness of the algorithm.2)In order to further improve the detection accuracy of the Cascade RCNN model,this paper uses data augmentation techniques such as Mixup to expand the sample size of the logo data set,so as to improve the sample diversity and balance the sample size of each category.At the same time,SoftNMS algorithm is used to optimize the screening mechanism of regional candidate frames,CIoU loss function is selected to improve the regression accuracy of prediction frames,and finally Adam W is used as the optimization function to improve the learning rate to enhance the recognition effect of the model.3)In the training process,the method of Transfer Learning is combined to speed up the convergence of the model by using the prior knowledge of the large data set.In addition,in order to prevent overfitting and improve the robustness of the algorithm,this paper adopts methods such as Weighted Random Sampler,Stochastic Weight Averaging and learning rate warmup to adjust the algorithm training process and make the algorithm model more satisfying to the needs of the training process.Finally,the improved logo detection algorithm is compared with several commonly used object detection algorithm frameworks.The experimental results show that the research algorithm in this paper has obvious advantages over the current popular object detection algorithms,and the m AP value of the improved Cascade R-CNN algorithm framework is improved by 4.5% compared with the unimproved algorithm framework.