Logo Recognition Research Based On Improved SURF Features

Logo recognition technology has a wide range of applications,such as recognizing the goods Logo to classify goods automatically,analyzing the brand influence by detecting Logo from network images,analyzing enterprise product ranking by recognizing Logo from e-commerce website.Domestic and overseas scholars have done related researches,but the recognition rate of Logo images in the network and natural scene needs to be improved.To solve the influence of Logo image and background environment diversity on the recognition rate,this paper proposes a Logo recognition method based on improved SURF features.The main work of this paper is as follows:(1)Aiming at the problem of Logo image scale changes,rotation changes,occlusion and intensity changes,we choose local invariant feature to describe Logo images.Considering the efficiency of SURF features,we analyze the SURF descriptor method to improve the SURF descriptor.The square neighborhood of the traditional SURF descriptor is changed into a circular neighborhood to enhance the rotation invariance of the descriptor and reduce the accumulated error of the descriptor.Then,the dimension of the descriptor is reduced to 32 dimensions,which improves the efficiency of the algorithm.(2)To solve the problem of high false matching rate caused by the complex background form network and natural scene in Logo images,we propose a novel matching algorithm based on the geometric position of the feature points.The experiment results show that the geometric position is sensitive to the rotation changes,so the angle of the test image needs to be corrected to eliminate the rotation difference between the test image and the sample image.Angle correction uses the main orientation of the SURF keypoints,and the main orientations of all the keypoints of the sample image and the test image are calculated respectively to find the highest frequency of the main orientations as the image representation orientation.Comparing the representation orientation of the sample image and the test image,the angle of the test image is compensated according to the representation orientation difference.The improved matching algorithm first calculates the matching angles between matching images,and then analyzes the distribution rule of correct matching angles and false matching angles.The adaptive clustering algorithm is used to cluster the matching angles,and the two steps filtering method combined with the number of elements in clusters and the number of clusters is used to eliminate the false matching angle.The experiments results show that the matching algorithm has high matching accuracy and low false rejection rate,which is beneficial to improve the accuracy of Logo image recognition.(3)In this paper,the algorithm is tested on the self-built drink Logo dataset and the public Flickr dataset.Compared with the related literature algorithms,the recognition rate of the proposed algorithm is improved by 13% and 5% respectively.