Vehicle Logo Detection Method Based On Deep Learning

Intelligent transportation system is more and more widely used in daily life.Correct identification of vehicle information is the key technology in intelligent transportation system.As an important part of the vehicle,logos can provide supplementary information for the detection n of vehicle information.The detection of the vehicle logo can help to promote the development of intelligent transportation systems.Moreover,the proportion of vehicle logos of the image is very small.It’s of great significance for the study of small object detection in a complex background.However,there are many problems in the current vehicle logo detection task.Such as difficulty in extracting small target features,diverse target shapes,and complex target backgrounds,which greatly increase the difficulty of detection.In order to improve the accuracy of vehicle logo detection,this paper will study from two aspects.Aiming at the problems of small vehicle logo target,complex shape and complex background information,a vehicle logo target detection method based on improved YOLOv4 model is proposed.Firstly,the CSPDense Net structure is used to improve the backbone feature extraction network and reconstruct the feature extraction network.The combination of dense connection structure and residual structure realizes the repeated extraction of small target features.Then,in order to solve the problem that the irregular shape of the vehicle logo affects the detection accuracy.The deformable convolution residual block is introduced to reconstruct the neck structure.At the same time,the neck network introduces the ECA-Net attention mechanism after upsampling to enhance the network ’s attention to vehicle logos.Finally,using Convolutional Transformer Block,a new detection head CT-Head is proposed.The CT-Head structure combines the local features of the vehicle logo and the global features of the vehicle to reduce the influence of the complex background on the vehicle logos detection.The mean average precision of the original YOLOv4 model on the large vehicle logo data set VLD-45 is 57.22 %,and the recall is 62.05 %.After the improvement of the model,the mean average precision is 62.94 %,increased by 5.72 %,the recall is 62.05 %,increased by 5.58 %.In order to further extract vehicle logo information from complex background and improve the accuracy of vehicle logo detection,a vehicle logo target detection method based on improved YOLOX model is proposed.Firstly,Swin Transformer Block is introduced to replace part of the CSP structure to improve the backbone feature extraction network.This method can make the model make better use of global features and reduce the interference of image background information on vehicle sign detection.Secondly,CSPAC module is proposed based on asymmetric convolution structure.The CSPAC module is used to replace the CSP structure of the original network neck,so as to highlight the local key features and prevent the loss of vehicle logo information.Finally,the Bicat feature fusion method is used to enhance the learning of channel features with stronger expression ability.Focus on vehicle logo information and strengthen the network ’s learning of vehicle logo features.The meanaverage precision of the original YOLOX model on the large vehicle logo dataset VLD-45 is 58.35 %.The mean average precision of the improved detection model on this data set is 64.21 %,an increase of 5.86 %.The experimental results show that the detection accuracy of the above two improved models on the large vehicle logo dataset VLD-45 has been improved to a certain extent.