Research On Vehicle Logo Recognition Based On Machine Learning

Vehicle-logo recognition can obtain the brand information of vehicle,and target detection is mainly used to solve the comprehensive problem of image location and classification.Therefore,this paper combines target detection with vehicle logo recognition,so as to quickly identify the information of vehicle logo.However,the traditional target detection algorithm needs to manually extract features,the training process is cumbersome,the detection speed is slow,and most of the feature extraction is not good enough shallow neural networks.YOLOv4 is a rapid one-stage target detection algorithm with high detection accuracy and speed.Therefore,this paper analyzes the relevant principles,methods and evaluation indicators of convolutional neural network and target detection technology by describing the research status of vehicle-logo recognition,and combining the application scenarios of vehicle-logo recognition.The input end,multi-scale feature fusion network and output end of YOLOv4 algorithm are improved,and the vehicle-logo detection algorithm VDNet is proposed.The main research and improvement work is as follows:(1)At the input end,K-means++ is used to replace the original K-means as the new calculation method of anchor box by analyzing the mechanism of anchor box,so as to realize more accurate selection of anchor box.At the same time,an adaptive image enhancement strategy is designed to optimize the training strategy and improve the detection rate.(2)In view of the multi-scale problem in detection,the original PANet is improved and the Bi FPN based vehicle-logo feature fusion network is proposed.The concept of weight is introduced to enable the network to integrate more features without increasing the cost basically,so that the network has better detection effect on targets of all scales.(3)At the output end,on the one hand,EIOU Loss is introduced as the new Loss function based on the analysis of the original Loss function CIOU Loss,so as to optimize the sample imbalance problem in the bounding box regression task;On the other hand,the weighted NMS method was adopted,and DIOU was used to replace the original IOU as the new weighting factor to optimize the elimination strategy of redundant boxes,which improved the accuracy of detection.In this paper,based on the improved VDNet algorithm,multiple groups of experiments were carried out by creating the data set of the Vehicle-logo and based on the classical algorithm,the original algorithm and the improved algorithm.The experimental results show that the VDNet algorithm has improved the training speed,detection speed and accuracy in the badge recognition,and has achieved a good balance between speed and accuracy.Finally,based on VDNET and the analysis of the requirements of the identification of the logo,the design and development of the logo identification software,the software can be real-time badge information acquisition,and deployed to other traffic systems,has a strong practical significance.