| With the continuous development of computer vision and Internet technology,logo detection technology is more and more widely used in real life such as commercial publicity,trademark infringement detection and urban intelligent transportation.In the field of commercial publicity,mining the logo information from pictures for brand tracking,and analyzing users’ preferences for brands to make personalized recommendations can greatly promote brand publicity.In the field of trademark infringement detection,intelligent logo images retrieval and identification can determine whether the applied trademark is infringed or the product is faked,so as to provide protection for intellectual property protection and stable development of society.In the field of urban intelligent transportation,due to the increase of vehicles,road congestion,frequent traffic accidents and other problems,road sign and license plate detection is very important to traffic safety.Therefore,logo detection is a meaningful research topic.For an image,logo detection is the process of finding the area where the logo is located in the image,and determining the specific category of the logo.Traditional logo detection methods have poor detection performance due to the limitations of manual feature extraction.At present,with the rapid development and application of convolutional neural networks,models based on deep learning show better detection performance.Therefore,according to the characteristics of logo images in real scenes,this work proposes a strategy based on the deep learning for logo detection.Specifically,by analyzing the characteristics of logo images and the limitations of current detection methods in logo detection tasks,it improves the basic methods to obtain better detection performance.The contribution of this work mainly includes the following two aspects:(1)A logo detection method based on deformed convolution and feature fusion is designed.When modeling logo,convolutional neural network has some limitations in detection performance because its fixed geometric structure cannot adapt to the geometric changes of logo.Aiming at the problem of poor detection performance caused by geometric changes,we propose a logo detection method based on deformed convolution and feature fusion.Specifically,an adjustable deformable convolution module is firstly added to the Res Net50 network to learn the offset and adjustment weight,so that the sampling position of the convolution kernel is adaptively adjusted according to the size and shape of the logo,and fully extracts different geometric changes.Then,a perceptual content feature reorganization operator is introduced into the feature pyramid network to generate an adaptive kernel at each position according to the semantic information of the feature map and perform feature reorganization to obtain a feature map containing more semantic information.The feature pyramid network integrates features of different scales,which significantly improves the effect of logo detection.Finally,the DIoU loss function is introduced in the regression process to achieve more accurate bounding box positioning.(2)A logo detection method based on dynamically trained classifiers and regressors is designed.During the training process,fixed settings in the network will limit the performance of the detector,such as the label assignment strategy(fixed IoU threshold setting)and fixed regression loss function,which cannot flexibly adapt to changes in the distribution of candidate boxes.Theoretically,high IoU threshold can obtain higher-quality candidate boxes,but setting a high IoU threshold at the beginning of training will lead to insufficient number of positive samples,while a low IoU cannot train a classifier to discriminate high-quality candidate boxes.In addition,in the regression branch,with the progress of network training,the distribution of regression labels also changes.As the quality of the candidate boxes improves,its mean and variance gradually decrease.The fixed setting reduces the contribution of high-quality candidate boxes to training.To solve the above problems,this work introduces a dynamic label assignment strategy and a dynamic smoothing loss function.Specifically,during the training process,the IoU threshold is automatically updated according to the distribution of candidate boxes,and higher-quality candidate boxes are gradually obtained.In the regression branch,the regression loss function parameters is adjusted,so that it automatically adjusts parameters to fit the distribution of regression labels,enhancing the contribution of high-quality candidate boxes to network training.In addition,this work uses convolutions with different dilation rates in the Res Net50 network,and adaptively selects the receptive field by using a switch function to detect logos with different scales.The two above proposed methods are experimentally evaluated on datasets of different scales.Firstly,a large number of comparative experiments are conducted with some methods with higher detection performance,and then the effectiveness of each module method is proved by ablation experiments.Finally,the visualization results are shown to illustrate that the method in this work has high detection performance for logo images in real scenes. |
