LOGO Classification And Detection Based On Deep Learning

With the development of the computer industry,image classification and object detection technologies have shown advantages in many fields such as school education,urban transportation,medical and health,consumer entertainment and so on.The researchers' visual processing of the massive commodity data in the e-commerce platform provides a different application environment for the logo classification and detection technology.Common applications are identifying counterfeit,fake and inferior products by identifying the logo in the image to achieve resolution detection;combining social media and logo image classification achieve commercial advertising preference analysis and personalized product recommendation;the detection of logo images realizes the determination of road signs and the construction of intelligent transportation systems.Traditional logo classification and detection techniques mostly use a single region of the image in a small dataset as the research point.However,its performance is low due to its simple recognition environment.Therefore,based on the characteristics of logo images and deep learning algorithms,the article expands the classification and detection of logo images,which mainly includes three aspects.(1)Construct datasets.There is a huge amount of logo image data in the real life,but most of the public logo datasets contain noise,which has become the bottleneck for deep learning algorithms.Therefore,this project has constructed Logo-2K+ and Logo Det-3K dataset through four steps of collecting,cleaning,labeling and analyzing.Meanwhile,we make these datasets public for use by the majority of peers.(2)Fine-grained classification algorithm for Logos.Due to the excessive number of categories,the logo dataset presents small differences among categories.The dataset has characters of multiple scales of objects,occlusion and background interference,etc.,making the classification more complicated.We proposed a Discriminative Region Navigation and Augmentation Network for logo classification,it used self-supervised training method that can locate logo regions of different scales in images with complex background only by the label information,and then perform feature enhancement via the guidance of regional features,and finally build a multi-feature set to achieve more accurate logo classification.(3)Design detection algorithm for logos.Due to the diversity of logo image scenes,the richness of trademark categories,the difficulty of processing massive product data,and the imbalance of data samples,logo detection is a great challenge.Using YOLOv3 as the basic framework to solve the problem of multi-scale targets in the logo image.Specifically,we use the K-means clustering method to recalculate the anchor selection of Logo Det-3K,and the Focal loss is solved the sample imbalance.At the same time,the CIoU regression loss is utilized to generate more accurate regression results.In summary,this paper conducts extensive experiments on the large-scale logo datasets(Logo-2K+ and Logo Det-3K),and the public datasets(Openlogo and Flicker-32).The experiments conducted fully demonstrate the feasibility of our proposed deep learning-based logo image classification and detection technology,which can contribute to the research of the logo images related algorithm.