Research And Implementation Of Real Scene Logo Detection Method Based On Deep Learning

With the rapid development of e-commerce,logo detection has shown great commercial value in the fields of product image retrieval,sports event sponsorship,marketing activity evaluation,and intellectual property protection.However,the diversity of logo categories,styles,and positions in real scenes poses a great challenge to logo detection technology.Deep learning provides new ideas for research on logo detection.With the amount of logo image data,deep learning technology allows researchers to quickly detect the position and category of the logo by only designing appropriate algorithm models.This thesis proposes a real scene logo detection method based on deep learning.First,filter and extract the most relevant areas of the logo instance in the image as the logo context to train the logo context discriminating network,and then generate suitable positions of logos in an image.Combined with the logo feature optimization scheme,it can synthesize real scene logo images as training data to train a multi-scale region proposal logo detection model.And through this way,the method can detect logos in real scenes.The logo detection method designed in this thesis includes five steps:logo context area selection,real scene logo image synthesis,multi-scale region proposal logo detection algorithm design,logo detection model training,and analysis of results.Among them,the key technologies are logo image synthesis and logo detection algorithm.This thesis proposes a data augmentation method based on logo context selection.By training the logo context discrimination network,this method can automatically determine the location where it is suitable for a logo in an image.Combined with the logo feature optimization scheme,it can quickly synthesize pictures,which solves the problems of insufficient data of public logo datasets and misleading of existing image synthesis methods.In the aspect of model design,this thesis proposes a multi-scale region proposal logo detection algorithm,which uses the ResNet convolutional structure to optimize image features based on Faster RCNN.At the same time,the feature pyramid multi-scale region proposal mechanism and bilinear interpolation algorithm are used to improve small-scale logo instances detection precision.The improved algorithm in the case of using only public dataset images and augmenting the training set with synthetic images obtained the AP@0.5 values of 85.8%and 90.3%respectively.Finally,the improved model is applied to the logo detection system,which proves that the real scene logo detection method designed in this thesis is applicable.