Research On Key Technologies Of Vehicle Recognition System Based On Deep Learning

With the improvement of national economic level and the improvement of infrastructure construction,the number of road vehicles continues to grow,which increases the management difficulty of traffic management departments.Vehicle attribute recognition system based on deep learning can be used for analyzing image intelligently and obtaining vehicles information from surveillance images.Obtaining vehicle information quickly and accurately can ease the management pressure of relevant departments and improve their work efficiency.In this paper,according to the functional requirements of the system,the relevant content is studied,and the following four functional modules are designed and realized.In order to solve the positioning problem of vehicles in the image,a vehicle detection method based on YOLOV4 is designed to complete the positioning task of vehicles in the image and classify vehicle types.YOLOV4 is a widely used target detection technology.Experiments show that YOLOV4 has high accuracy and fast reasoning speed.The Retinaface model used for face task is modified,so that it not only realizes the function of license plate location and four corners detection,but also recognizes the type of license plate.After the key information of license plate is obtained,the license plate image is aligned and corrected by perspective transformation to reduce the difficulty of character recognition.LPRNET is a lightweight end-to-end character recognition model that uses a full convolutional network to align the character recognition tasks of license plate images.This paper improves the mean operation in the reasoning process of LPRNET,and the experimental results show that the improved LPRNET single license plate recognition accuracy is greatly improved.A vehicle multi-attribute recognition technology using attention module and weakly supervised data augmentation is designed,and it can complete the identification task of vehicle make,color and attitude attributes.Experimental results show that the introduction of attention and weak supervision data augmentation mechanism can greatly improve the accuracy of vehicle multi-attribute recognition.The vehicle attribute recognition system is realized by using the above mentioned technology,and the system achieves vehicle structured information acquisition in road monitoring images.