Trademark Detection In Natural Scene Base On Synthetic Data

In the field of computer vision,trademark recognition is a highly concerned issue,which is widely used in copyright monitoring,advertising,precise marketing and other fields.If trademark recognition can be well identified,many applications with commercial value will be developed accordingly.However,in view of lack of human-labelled data in trademark detection and the problem of small object of trademarks in images,it is difficult to trademark recognition.This paper aims to recognize the trademark in book image and provide technical basis for copyright monitoring.It improves the precision of trademark detection from the following two aspects.(1)A method of synthesizing data containing specific markers is proposed,and the marking files are automatically generated during computer synthesis,which realizes the rapid acquisition of a large number of marking data.The Canny operator is used to remove the background of the trademark.Then,Affine Transformation is applied to simulate the real situation of trademark deformation.Linear mapping is presented by matrix multiplication and translation is represented by vector addition.The simulation of trademark deformation and deformation in original data is realized.At last,bilinear interpolation is used to integrate the trademark into the image of the book.In addition,randomly adds additional interference information,so that the probability distribution of the synthetic data is closer to the probability distribution of the real data.The average accuracy of the original data set is increased by 1.6%and that of the synthetic data set is increased by 26.1%.It solves the difficulty of obtaining trademark images and the high cost of image annotation.(2)The method based on multi-strategy fusion is studied to improve the accuracy of small object detection.Firstly,through Feature Pyramid Networks,adds up sampling and pixel values,so that low-level texture f-eatures of the image extracted by ResNet is combined with high-level semantic features.It reduces the serious loss of location information of small objects after multiple sampling.Secondly,RoIAlign cancels quantization operation to make the local feature mapping more accurate,which improves the accuracy of feature mapping of small target objects.Finally,the hard example which are misclassified but have high confidence are mined.It integrates the difficulty feedback of the entropy as a sample into the loss,increases the contribution of the difficult sample to the model,and improves the judgment of the model for the difficult sample such as small object.By combining various optimization methods,the average accuracy of the model is improved by 74.7%,which alleviates the difficulty of small target detection.After optimization,the mAP in the original book picture set was increased from 0.3586 to 0.6366,and mAP in the synthetic book picture set was increased from 0.2975 to 0.6069.This shows that the model has strong generalization ability and the synthetic data effect is close to the real data.The trademark detection framework is integrated into the digital image trademark detection and monitoring system,which can automatically detect and identify the image trademark according to the user's needs,and monitor the books and trademarks on the e-commerce platform.It has been put into use in the copyright monitoring of the textbook of the 2018 Junior Accounting Title Examination of the Economic and Financial Publishing House,and has high academic value and use value.