Research On Trademark Detection Algorithm Based On Deep Learning

In the field of computer vision,the performance of deep learning in target detection tasks has surpassed the past and reached the highest level in history.The main research topic of this paper is the application of target detection in the field of trademarks.Different from the target detection task in the open field,the trademark detection task has the characteristics of large target scale span,complex scene change,serious target rotation deformation and large number of small targets.However,the current mainstream target recognition algorithm is for the recognition task in the trademark field.The accuracy is not high.Based on the above background,the main content of this thesis is the research of target detection algorithm in the field of trademark detection.Aiming at the problems of small target,occlusion and rotation in the trademark dataset,this paper proposes a candidate region generation algorithm based on multi-scale image template.The main idea of the algorithm is to generate a network framework based on the original candidate frame,and combine the image pyramid and multi-scale image template into the unified network.At the same time,the feature advance network is changed from the original last layer feature to the feature pyramid network,which improves the algorithm to small.Target detection capability.For different resolution images,the algorithm uses image templates of different size scales to solve the problem of multi-scale detection and no excessive redundancy calculation.Through the comparison experiments on the public dataset and the manually labeled car domain dataset CarLogos,we prove that our algorithm is better than the original candidate frame extraction algorithm and the Selective Search algorithm.Because the candidate region generation algorithm proposed in the previous paper can not input multi-scale images,the algorithm can not be unified into the original Faster R-CNN framework.This paper introduces the idea of full convolutional neural network,and replaces the original with a convolutional layer of 1*1 convolution kernel size.A fully connected layer,a trademark detection algorithm based on full convolutional neural network is proposed.Based on the Faster R-CNN framework,the algorithm combines a full convolutional neural network with a candidate region generation algorithm based on multi-scale image templates to integrate testing and training in a unified network.Through experimental analysis,the algorithm has better mAP values on the FlickrLogos dataset and CarLogos dataset than the original Faster R-CNN algorithm and several other mainstream target detection algorithms.In view of the fact that there are few dataset categories and single scenarios in the current trademark detection data collection,this paper also provides a trademark dataset for the automotive field.This article uses the web crawler to crawl a large number of trademark images of the automobile field under natural scenes and manually label them.There are many categories of datasets,and each category has a large number of labels,which can effectively enrich the trademark detection dataset.