Research On Logo Detection And Recognition Algorithm Based On Synthetic Data

With the rapid development of deep learning and computer vision technology in recent years,the size of Internet pictures and video data is getting larger and larger,and the information in images plays an increasingly important role,especially detection of logos in pictures and videos,is a very important issue.Logo detection has many useful applications in real life,such as business analysis of brands and brand infringement.Logo as a unique symbol of a company,organization or product,merchants can search for relevant logos to analyze the development of their brands in the entire market and future trends,and at the same time help advertisers to check the effectiveness of advertising and if there is a problem with copyright infringement.The vehicle-logo detection can help to achieve intelligent transportation systems,even AR.A logo can be either a text symbol,a graphic,or a mixture of the two.At present,the main difficulties in logo detection include the position and context of the logo in the picture is uncertain.Due to various lighting effects,occlusion,rotation,cutting effect and scale of the natural scene,the logo has changed greatly and the intra-class differences of the logos are relatively large,and some differences between classes may be relatively small,which may lead to false detections;vehicle logos generally have problems with smaller goals;deep network models require a large amount of annotated data,and current mainstream logos detected publicly annotated data set has the problems of few logos and less label data,which is not conducive to the training of the model.Aiming at the problem of insufficient training data,this paper proposes a logo data synthesis method based on background segmentation.Since the strategy of synthesizing logos directly on the background image does not take into account changes in the background,when the logos are randomly synthesized,especially text-based logos,in this complex context,the synthetic logos may appear unrealistic.There are effects that can interfere with model learning.This paper generates a more realistic sample by segmenting the background image,processing the logo template based on the image processing method,and processing at the final synthesisIn view of the diversity and indistinguishability of logos in natural scenes,this paper synthesizes more abundant and diverse data through data synthesis method.On the other hand,in terms of algorithm models,this paper introduces two dimensions in the feature extraction stage.Attention mechanism increases the weight of features with strong discrimination and suppresses the feature of weak discrimination.At the same time,this paper introduces similarity learning in the algorithm,and further enhances the classification ability of the model to the target by adding the triplet loss to the loss function.In view of the problem of logo occlusion in natural scenes,this paper introduces multi-region context information in the algorithm,which increases the detection and recognition ability of logos under occlusionAiming at the problem that there are a lot of small targets and uneven size distribution in the natural scene,this paper uses the multi-scale feature extraction network to extract features,and uses the Kmeans algorithm based on IoU distance to cluster the logo size.In the detection and identification stage,it is introduced.Target context information classifies targets to improve the accuracy of small target detectionIn the last part of the experiment,the validity of the data synthesis algorithm and the detection and recognition algorithm are verified respectively,and the advantages of this paper are verified by comparison with related algorithms.