Research On Logo Detection Algorithm Based On Deep Learning

With the rapid development of mobile Internet,the image in the internet has a great growth.the massive information in these images is very useful,such as brand analysis for users and social media can also analyse brand preference of uses to do personalized recommendation.Detecting logo in massive images is used to mining brand infromation in this paper.Although in recent years,the method of object detection based on deep learning has achieved great success,but the logo detection in a image is not improved as well,the main problem is the existence of multi-scale logo and rotated logos.Not noly that,logo is usually quite small compare to usual objects.In order to solve these three problems,this paper proposed a multi-scale region proposal network and context-based classification network to solve the problem existed in Logo detection,the primary research results are as follows:Aiming at the problem of direction rotated logo object,we introduced a adaptive convolution layer and adaptive pooling layer to extract the feature of image and it improves the robust ability of our system.For the problem of multi-scale detection in logo detection,an improved multi-scale region proposal network is proposed.Different from the original region proposal network,the network uses the method of feature pyramid to achieve multi-scale target detection.Not only that,according to the characteristics of the common logo image size relative to the target is relatively small,this paper uses the k-means algorithm based on Io U distance to do scale clustering in logo objects and obtain the distribution of logo size.Therefore,we can generate semantic feature pyramid with convolution network and generate region proposal in different scales by using semantic feature pyramids.A comparison experiment is conducted to verify the improved multi-scale region proposal network,compared with the original candidate region extraction network,the recall rate of region proposals extracted from a variety of data sets is greatly improved.Aiming at the problem that the target in logo detection is relatively small and difficult to recognize,this paper proposes a classification algorithm based on context-based classification network and long short-term memory model to classify logo targets.The context-based classification network uses the contextual information of the target when classifying the target.Different scales of contextual information of the target are used as the inputs of long short-term memory model to get the classification results.Long short-term memory model makes efficient use of different scales of contextual information to do classification.Finally,in order to improve the accuracy of the bounding box,the method of small bounding box regression is used and improved the detection accuracy.Compared with the traditional object detection methods,the performance of context-based detection network on the image Logo-50 dataset is improved 6 percentage points m AP compared with the traditional object detection methods.