Logo Detection And Recognition In Images On Social Networking

In recent years, with the development and spreading of social networks, peoplespend more and more time on social networks, which makes social networksbecome a advertising and business platform potentially. There is Brand trackingarising as a service recently. The popularity of brand can be evaluated by analyzingthe brand propagation force and the customer review in the media. Because of thedevelopment of social networks, brand tracking has been transferred to socialnetworks gradually.For brand tracking, only keyword search function is provided in the socialnetworking platforms currently, which brings two drawbacks: first, despitecontaining keywords, a lot of noise is included in search results, but they often isnot relevant with the brand; second, a large amount of information containing on thebrand image can not be retrieved. In order to solve the second problem, a newLOGO(trademark) detection method which is detecting the LOGO in the image thatis uploaded by social network user is proposed in this paper. This method not onlycan be used as a standalone application, brand tracking function directly but also beused as a part of the brand analysis system.Most of images on social network is taked by user. So the quality of image isoften low, including the poor lighting conditions, image blurring, the shooting angledifference, which makes the image of the LOGO asymmetry illumination, tiltrotation, elastic deformatio, partially occluded and other issues. Furthermore, inorder to make LOGO recognizable, LOGO is often designed as a simple appearence,which makes LOGO similar to other images. These have increased the difficulty ofdetecting LOGO. In order to solve the social network LOGO image detectionproblem, LOGO detection methods based on machine learning is proposed in thispaper. The authors have evaluate its application on social networks.There are three contributions in the paper. On the one hand, building a trainingset and a test set which contain100kinds of LOGO images. The training set includethe LOGO position, size and angle of rotation of training samples. In test setincluding one million images from Sina Weibo, each image has been marked good ifit contains LOGO. If LOGO is cotained in one test sample, the location and size isalso recorded. In training set the number of samples per LOGO average is over300.The data set covers the LOGO in different lighting, plane rotation, blur, shootingangle of the case. So the training set have great value and tested by researchers. Theother hand, a LOGO detection algorithm is proposed based on new AdaBoost. Asmachine learning methods is used for LOGO detection in this paper, this is a problem which is serious asymmetric between the positive and negative sample.AdaBoost need to specify the process of training the proportion of positive andnegative samples. So features each AdaBoost node selection out as input result inlinear classifier.Finally, this paper propose a new detection algorithm based onLOGO brand tracking method, by determining whether the image contains LOGOin a social network to evaluate the popularity of brand and analysize the brand bystages, which supplement the existing text-based keyword function. This methodcan be used as the automatic method without manual intervention.