Study On Vehicle Logo Recognition Based On Deep Learning And Transfer Learning

As one of the key features of a vehicle,the vehicle logo can be combined with other vehicle feature recognition such as license plate recognition and vehicle color recognition,which can greatly improve the accuracy of vehicle recognition.In many applications such as traffic security checkpoints,vehicle illegal decks,unattended parking lots,etc.,vehicle identification will play a more important role in obtaining complete information about the vehicle.The complete vehicle logo recognition process covers vehicle detection,vehicle positioning,and vehicle identification.The research on vehicle detection is relatively mature,but the speed and accuracy of vehicle detection need to be improved.However,the current research on the positioning of vehicle logos is still incomplete.Most of the current researches focus on the image of the vehicle’s forward viewing angle and the symmetry of the license plate area as the research object.Vehicle positioning problem in complex scenes such as,tilt,etc.The article proposes a method of vehicle logo positioning based on the coordinates of the vertex coordinates of the license plate,so as to realize the vehicle logo positioning technology.Based on the positioning of the vehicle logo,the design of the vehicle logo recognition network based on the YOLOv3 network is carried out,and then a complete automatic vehicle logo recognition system is researched on the realization of the project.From the perspective of building an automatic vehicle logo recognition system,the research contents of this paper mainly include:(1)Fast and accurate vehicle detection.The actual input image usually contains multiple targets,so it is necessary to detect the vehicle.Because the YOLOv3 network runs fast and has multi-scale detection capabilities,this article will migrate the YOLOv3 network and its weights trained in the COCO dataset,and only extract vehicle images by processing the output prediction categories.After the vehicle image is extracted,the vehicle image is enlarged in size in combination with the existing methods to facilitate subsequent image processing.(2)Vehicle logo positioning based on license plate position.Vehicle logo positioning is the key technology of the automatic vehicle logo recognition system.The article first locates the four apex coordinates of the license plate and analyzes the relative position relationship between the license plate and the license plate.A method is proposed to obtain the area above the license plate that contains the license plate using coordinate transformation,and then extract Contains car logo images for car logo positioning in complex application scenarios.Finally,according to the coordinates of the four vertices of the vehicle logo area and the coordinates of the vertices of the target area,perspective transformation is performed to correct the vehicle logo image.The vehicle logo image processed by perspective transformation has approximately a positive perspective and a fixed size,and prepares data for subsequent vehicle logo recognition.(3)Design of vehicle logo recognition network and construction of automatic vehicle logo recognition system.This part will label the fixed-size car logo images to build the car logo training data set.According to the size characteristics of the car logo image,a network design of car logo recognition based on YOLOv3 network is performed,and training is performed.Finally,the three modules of vehicle detection,vehicle positioning,and vehicle identification are integrated under the Ubuntu system to form a complete automatic vehicle identification system.Generally speaking,The article’s research on vehicle logo recognition is mainly divided into three modules: vehicle detection,vehicle logo positioning,vehicle logo recognition,and an overall automatic vehicle logo recognition system.Experiments show that the vehicle detection module has a strong generalization ability,and the detection accuracy of the vehicle is more than 98.4%;for the vehicle image with the position relationship of the vehicle logo directly above the license plate,the accuracy of the vehicle logo positioning is more than 96.7%;The detection module,labeled data is trained according to the designed vehicle logo recognition network.Finally,the complete vehicle logo recognition system is verified,and the overall vehicle logo recognition rate reaches 88.3%.